J. Fluid Mech. (2015), vol. 774, pp. 395–415.

c Cambridge University Press 2015 395

doi:10.1017/jfm.2015.268

Direct numerical simulation of turbulent

channel flow up to Re τ ≈ 5200

Myoungkyu Lee1 and Robert D. Moser1,2, †

1 Department of Mechanical Engineering, The University of Texas at Austin, TX 78712, USA
2 Center for Predictive Engineering and Computational Sciences, Institute for Computational Engineering
and Sciences, The University of Texas at Austin, TX 78712, USA

(Received 18 September 2014; revised 20 March 2015; accepted 6 May 2015;

first published online 10 June 2015)

A direct numerical simulation of incompressible channel flow at a friction Reynolds

number (Reτ ) of 5186 has been performed, and the flow exhibits a number of the
characteristics of high-Reynolds-number wall-bounded turbulent flows. For example,
a region where the mean velocity has a logarithmic variation is observed, with von
Kármán constant κ = 0.384 ± 0.004. There is also a logarithmic dependence of the
variance of the spanwise velocity component, though not the streamwise component.
A distinct separation of scales exists between the large outer-layer structures and small
inner-layer structures. At intermediate distances from the wall, the one-dimensional
spectrum of the streamwise velocity fluctuation in both the streamwise and spanwise
directions exhibits k−1 dependence over a short range in wavenumber (k). Further,
consistent with previous experimental observations, when these spectra are multiplied
by k (premultiplied spectra), they have a bimodal structure with local peaks located
at wavenumbers on either side of the k−1 range.
Key words: turbulence simulation, turbulent boundary layers, turbulent flows

1. Introduction
Recently, relatively ‘high’-Reynolds-number wall-bounded turbulence has been
investigated with the help of innovations in measurement technologies (Nagib et al.
2004; Kunkel & Marusic 2006; Westerweel, Elsinga & Adrian 2013; Bailey et al.
2014) and computation (Borrell, Sillero & Jiménez 2013; El Khoury et al. 2013;
Lee, Malaya & Moser 2013). Because of the simplicity in geometry and boundary
conditions, pressure-driven turbulent flow between parallel walls (channel flow) is
an excellent vehicle for the study of wall-bounded turbulence via direct numerical
simulation (DNS). Since Kim, Moin & Moser (1987) showed agreement between
DNS and experiments in channel flow in 1987, DNS of channel flow has been used
to study wall-bounded turbulence at ever higher Reynolds numbers, as advances in
computing power have allowed. For example, channel flow DNS has been performed
at Reτ = 590 (Moser, Kim & Mansour 1999), Reτ = 950 (Del Álamo et al. 2004) and
Reτ = 2000 (Hoyas & Jiménez 2006). More recently, simulations with Reτ ≈ 4000
were performed separately by Lozano-Durán & Jiménez (2014) in a relatively small

† Email address for correspondence: [email protected]

396 M. Lee and R. D. Moser
domain size and by Bernardini, Pirozzoli & Orlandi (2014). Moreover, to investigate
the differences between channel flow turbulence and boundary layer turbulence, a
simulation at Reτ = 2000 of a zero-pressure-gradient boundary layer was performed
by Sillero, Jiménez & Moser (2013).
The idealized channel flow that is so straightforward to simulate is more difficult
to realize experimentally because of the need for sidewalls. Measurements of channel
flow are not as numerous as for other wall-bounded turbulent flows, but a number of
studies have been conducted over the years (Comte-Bellot 1963; Dean & Bradshaw
1976; Johansson & Alfredsson 1982; Wei & Willmarth 1989; Zanoun, Durst & Nagib
2003; Monty & Chong 2009; Zanoun, Nagib & Durst 2009). Of particular interest
here will be the channel measurements made by Schultz & Flack (2013) at Reynolds
number up to Reτ = 6000, because they bracket the Reynolds number simulated here.
Turbulent flows with Reτ of the order of 103 and greater are of interest because
this is the range of Reynolds numbers relevant to industrial applications (Smits
& Marusic 2013). Further, it is in this range that characteristics of wall-bounded
turbulence associated with high Reynolds number are first manifested (Marusic et al.
2010b), and so studies of these phenomena must necessarily be conducted at these
Reynolds numbers. The best-known feature of high-Reynolds-number wall-bounded
turbulence is the logarithmic law in the mean velocity, which has been known since
the 1930s (Millikan 1938). The von Kármán constant κ that appears in the log law
is also an important parameter for calibrating turbulence models (Spalart & Allmaras
1992; Durbin & Pettersson Reif 2010; Marusic et al. 2010b; Smits, McKeon &
Marusic 2011). It was considered universal in the past, but Nagib & Chauhan (2008)
showed that κ can be different in different flow geometries. Further, Jiménez &
Moser (2007) found that channel flows with Reτ up to 2000 do not exhibit a region
where the mean velocity profile strictly follows a logarithmic law, and showed that
a finite-Reynolds-number correction like that introduced by Afzal (1976) was more
consistent with data in this Reynolds number range. Similarly, Mizuno & Jiménez
(2011) used a different finite-Reynolds-number correction to represent the overlap
region in boundary layers, channels and pipes.
For turbulent intensities, Townsend (1976) predicted that, in high-Reynolds-number
flows, there are regions where the variance of the streamwise and spanwise velocity
components decreases logarithmically with distance from the wall. This has been
observed experimentally for streamwise (Kunkel & Marusic 2006; Hultmark et al.
2012; Hutchins et al. 2012; Winkel et al. 2012; Marusic et al. 2013) and spanwise
components (Fernholz & Finley 1996; Morrison et al. 2004), but has only been
observed in the spanwise component in DNS (Hoyas & Jiménez 2008; Sillero et al.
2013). Further, the peak of the streamwise velocity variance has a weak Reynolds
number dependence when it is scaled by the friction velocity, uτ (DeGraaff & Eaton
2000; Hoyas & Jiménez 2006).
There has also been recent interest in the role of large-scale motions (LSMs) in
high-Reynolds-number wall-bounded turbulent flow (Kim & Adrian 1999; Hutchins
& Marusic 2007; Wu, Baltzer & Adrian 2012). According to the scaling analysis of
Perry, Henbest & Chong (1986) there is a region where the one-dimensional spectral
energy density has a kx−1 dependence (kx , wavenumber in streamwise direction). This
has been supported by experimental evidence (Nickels et al. 2005, 2007; Dixit &
Ramesh 2013), but not numerical simulations.
In summary, there are many characteristics of high-Reynolds-number wall-bounded
turbulence that have been suggested by theoretical arguments and corroborated
experimentally, but which have not been observed in direct numerical simulations,
 DNS of turbulent channel flow up to Reτ ≈ 5200 397
presumably due to the limited Reynolds numbers of the simulations. This is
unfortunate because such simulations could provide detailed information about these
high-Reynolds-number phenomena that would not otherwise be available. The DNS
reported here was undertaken to address this issue, by simulating a channel flow at
sufficiently high Reynolds number and with a sufficiently large spatial domain to
exhibit characteristics of high-Reynolds-number turbulence like those discussed above.
The simulation was performed for Reτ ≈ 5200 with the same domain size as used by
Hoyas & Jiménez (2006) in their Reτ = 2000 simulation.
This paper is organized as follows. First, the simulation methods and parameters are
described in § 2, along with other simulations and experiments used for comparison.
Results of the simulation that arise due to the relatively high Reynolds number of the
simulation are presented in § 3. Finally, conclusions are offered in § 4.

2. Simulation details
In the discussion to follow, the streamwise, wall-normal and spanwise velocities will
be denoted u, v and w respectively, with the mean velocity indicated by a capital letter
and fluctuations by a prime. Furthermore, h·i indicates the expected value or average.
Thus, U = hui and u = U + u0 .
The simulations reported here are DNS of incompressible turbulent flow between
two parallel planes. Periodic boundary conditions are applied in the streamwise (x) and
spanwise (z) directions, and no-slip/no-penetration boundary conditions are applied at
the wall. The computational domain sizes are Lx = 8πδ and Lz = 3πδ, where δ is the
channel half-width, so the domain size in the wall-normal (y) direction is 2δ. The flow
is driven by a uniform pressure gradient, which varies in time to ensure that the mass
flux through the channel remains constant.
A Fourier–Galerkin method is used in the streamwise and spanwise directions, while
the wall-normal direction is represented using a B-spline collocation method (Kwok,
Moser & Jiménez 2001; Botella & Shariff 2003). The Navier–Stokes equations
are solved using the method of Kim et al. (1987), in which equations for the
wall-normal vorticity and the Laplacian of the wall-normal velocity are time-advanced.
This formulation has the advantage of satisfying the continuity constraint exactly
while eliminating the pressure. A low-storage implicit–explicit scheme (Spalart,
Moser & Rogers 1991) based on third-order Runge–Kutta for the nonlinear terms
and Crank–Nicolson for the viscous terms is used to advance in time. The time
step is adjusted to maintain an approximately constant Courant–Friedrichs–Lewy
(CFL) number of one. A new highly optimized code was developed to solve the
Navier–Stokes equations using these methods on advanced petascale computer
architectures. For more details about the code, the numerical methods and how
the simulations were run see Lee et al. (2013, 2014).
The simulations performed here were conducted with resolution comparable to that
used in previous high-Reynolds-number channel flow simulations, when measured in
wall units. Normalization in wall units, that is with kinematic viscosity ν and √
friction
velocity uτ , is indicated with a superscript ‘+’. The friction velocity is uτ = τw /ρ,
where τw is the mean wall shear stress and ρ is density. For the highest-Reynolds-
number simulation reported here, which is designated LM5200, Nx = 10 240 and Nz =
7680 Fourier modes were used to represent the streamwise and spanwise directions,
which results in an effective resolution of 1x+ = Lx+ /Nx = 12.7 and 1z+ = Lz+ /Nz =
6.4. In the wall-normal direction, the seventh-order B-splines are defined on a set
of 1530 = Ny − 6 knot points (Ny is the number of B-spline basis functions), which
398 M. Lee and R. D. Moser

Name Reτ Reb Method Lx /δ Lz /δ 1x+ 1z+ 1y+ +

w 1yc Ny Tuτ /δ Line and symbols
LM180 182 2 857 SB 8π 3π 4.5 3.1 0.074 3.4 192 31.9 (Magenta)
LM550 544 10 000 SB 8π 3π 8.9 5.0 0.019 4.5 384 13.6 (Blue)
LM1000 1000 20 000 SB 8π 3π 10.9 4.6 0.019 6.2 512 12.5 (Red)
LM5200 5186 125 000 SB 8π 3π 12.7 6.4 0.498 10.3 1536 7.80 (Black)
HJ2000 2003 43 650 SC 8π 3π 12.3 6.1 0.323 8.9 633 11 (Green)
LJ4200 4179 98 302 SC 2π π 12.8 6.4 0.314 10.7 1081 15 (Blue)
BPO4100 4079 95 667 FD 6π 2π 9.4 6.2 0.010 12.5 1024 8.54 (Red)
SF4000 4048 94 450 LDV ··· ··· ··· ··· ··· ··· ··· ···
SF6000 5895 143 200 LDV ··· ··· ··· ··· ··· ··· ··· ···
TABLE 1. Summary of simulation parameters. Here, 1x and 1z are in terms of
Fourier modes for spectral methods, 1yw and 1yc represent the grid spacing at the
wall and the centreline respectively, δ is the channel half-width, Reτ = uτ δ/ν, Tuτ /δ
is the total simulation time without transition; SB denotes spectral/B-spline, SC denotes
spectral/compact finite difference, FD denotes finite difference and LDV denotes laser
doppler velocimetry.

are distributed uniformly in a mapped coordinate ξ that is related to the wall-normal

coordinate y through

y sin(ηξ π/2)
= , −1 6 ξ 6 1. (2.1)
δ sin(ηπ/2)

The single parameter η in this mapping controls how strongly the knot points are
clustered near the wall, with the strongest clustering occurring when η = 1. In
LM5200, η = 0.97, resulting in the first knot point from the wall at 1y+ w = 0.498 and
the centreline knot spacing of 1y+c = 10.3. The N y collocation points are determined as
the Greville abscissae (Johnson 2005), in which, for n-degree splines, each collocation
point is the average of n consecutive knot points (n = 7 here), with the knots at the
boundary given a multiplicity of n − 1. As a result, the collocation points are
more clustered near the wall than the knot points. Resolution parameters for all the
simulations discussed here are provided in table 1.
Because the mass flux in the channel remains constant, the bulk Reynolds
number RebR can be specified directly for a simulation, where Reb = Ub δ/ν,
δ
Ub = (1/2δ) −δ U(y) dy and U(y) is the mean streamwise velocity. Four simulations
at four different Reynolds numbers were conducted. Of most interest here is the
highest-Reynolds-number case, LM5200, for which the bulk Reynolds number is
Reb = 1.25 × 105 and the friction Reynolds number is Reτ = 5186 (Reτ = uτ δ/ν).
The three other cases simulated, LM180, LM550 and LM1000, were performed
for convenience to regenerate data for previously simulated cases (Kim et al. 1987;
Moser et al. 1999; Del Álamo et al. 2004) using the numerical methods used here.
The simulation details for each case are summarized in table 1.
In addition, channel flow data from four other sources in the literature are included
here for comparison. The first is a simulation at Reτ = 2000 conducted by Hoyas &
Jiménez (2006, HJ2000), which used the same domain size and a similar numerical
scheme to the current simulations, differing only in the use of high-order compact
finite differences in the wall-normal direction, rather than B-splines. A second
simulation (LJ4200) by Lozano-Durán & Jiménez (2014) was at Reτ = 4179 and
 DNS of turbulent channel flow up to Reτ ≈ 5200 399

(a) 0.010 (b) 0.004

0.005 0.002
Residual

0 0

–0.005 –0.002
0 0.2 0.4 0.6 0.8 1.0 10 0 10 1 10 2 10 3 10 4

F IGURE 1. (Colour online) Statistical stationarity of simulation. (a) The residual in (2.2),
with the linestyle legend given in table 1; the dotted line is the standard deviation of
the estimated total stress in LM5200. (b) The residual in (2.3) of LM5200 (solid), and
the standard deviation of the estimated statistical error in the sum of the right-hand side
terms in (2.3) (dashed).

used the same numerical methods as HJ2000, but the domain size in x and z was
much smaller. The third simulation (BP04100), which was performed by Bernardini
et al. (2014), was at Reτ = 4079 and used a domain size not much smaller than that
used here, but these simulations were performed using second-order finite differences.
Finally, experimental data from laser Doppler velocimetry measurements at two
Reynolds numbers (Reτ = 4048 and 5895, SF4000 and SF6000 respectively) are
reported by Schultz & Flack (2013) and are also included here for comparison.
Summaries of all these data sources are also included in table 1.
We used the method described by Oliver et al. (2014) to estimate the uncertainty in
the statistics reported here due to sampling noise. For LM5200, the estimated standard
deviation of the mean velocity is less than 0.2 % and the estimated standard deviation
of the variance and covariance of the velocity components is less than 0.5 % in the
near-wall region (y+ < 100 say) and 3 % in the outer region (y > 0.2δ say). We also
used the total stress and the Reynolds stress transport equations to test whether the
simulated turbulence was statistically stationary. In a statistically stationary turbulent
channel, the total stress, which is the sum of Reynolds stress and mean viscous stress,
is linear due to momentum conservation:
∂U + y
+
− hu0 v 0 i+ ≈ 1 − . (2.2)
∂y δ
As shown in figure 1(a), the discrepancy between the analytic linear profile and total
stress profile from the simulations is less than 0.002 (in plus units) in all simulations,
and it is much smaller than the standard deviation of the estimated total stress of
LM5200.
The Reynolds stress transport equations, which govern the evolution of the Reynolds
stress tensor, are given by
Dhu0i u0j i ∂hu0i u0j u0k i ∂ 2 hu0i u0j i
 
∂Uj ∂Ui
= − hu0i u0k i + hu0j u0k i − +ν
Dt ∂xk ∂xk ∂xk ∂xk ∂xk
0  0 0   0 0
∂hp ui i ∂hp uj i
  0 0 0
∂ui ∂uj ∂ui ∂uj

0
+ p + − + − 2ν . (2.3)
∂xj ∂xi ∂xj ∂xi ∂xk ∂xk
400 M. Lee and R. D. Moser
While not reported here, all the terms on the right-hand side of (2.3) have
been computed from our simulations, and the data are available at http://turbulence.
ices.utexas.edu. In a statistically stationary channel flow, the substantial derivative on
the left-hand side of (2.3) is zero. Hence, any deviation from zero of the sum of the
terms on the right-hand side of (2.3) is an indicator that the flow is not stationary.
The residual of (2.3) is shown in figure 1(b) (solid lines) in wall units. The values
for all components of the Reynolds stress are much less than 0.0001 in wall units,
which is less than 0.01 % error in the balance near the wall. The relative error in the
balance increases to of the order of 1 % away from the wall, as the magnitude of the
terms in (2.3) decreases. Across the entire channel, the estimated standard deviation
of the statistical noise (dashed lines) is much larger than these discrepancies.

3. Results
3.1. Mean velocity profile
The multiscale character of wall-bounded turbulence, in which ν/uτ is the length
scale relevant to the near-wall flow and δ applies to the flow away from the wall, is
well known. As first noted by Millikan (1938), this scaling behaviour and asymptotic
matching lead to the logarithmic variation of mean streamwise velocity with the
distance from the wall in an overlap region between the inner and outer flows. This
‘log law’ is given by
1
U + = log y+ + B, (3.1)
κ
where κ is the von Kármán constant. In a log layer, the indicator function

∂U +
β(y+ ) = y+ (3.2)
∂y+
is constant and equal to 1/κ. Hence, the indicator function, β, will have a plateau if
there is a logarithmic layer.
The mean streamwise velocity profile is shown in figure 2 for all the data sets
listed in table 1. The profiles from all the relatively high-Reynolds-number cases
are consistent, as expected. Despite this agreement, the indicator function β shows
some disagreement between the three highest-Reynolds-number simulations, as shown
in figure 3(a). In the LM5200 case, β is approximately flat between y+ = 350 and
y/δ = 0.16 (y+ = 830), indicating a log layer in this region. The LJ4200 case also
appears to be converging towards a plateau in this region, but there is apparent
statistical noise in the profile, which is understandable given the small domain size.
However, the BPO4100 simulation does not have a plateau in β. There is also a small
discrepancy between BPO4100 and the other two cases from y+ ≈ 30 to 100. These
discrepancies are significantly larger than the statistical uncertainty in the value of
β (approximately 0.2 %) in the current simulations, and presumably in the BPO4100
simulations, as the averaging time and domain sizes are comparable.
The values of κ and B in (3.1) were determined by fitting the mean velocity data
from LM5200 in the region between y+ = 350 and y/δ = 0.16 to obtain the values of
0.384 ± 0.004 and 4.27 respectively, with R2 = 0.9999, where R2 is the coefficient of
determination, which is one for a perfect fit. The value of κ agrees with the value
computed by Lozano-Durán & Jiménez (2014), but shows a slight discrepancy with
the values measured by Monty (2005) and Nagib & Chauhan (2008), which are κ =
0.37 and 0.39 respectively. The value of κ obtained here is remarkably similar to those
 DNS of turbulent channel flow up to Reτ ≈ 5200 401

30
25
20
15
10
5
0
10 0 10 1 10 2 10 3 10 4

F IGURE 2. (Colour online) Mean streamwise velocity profile for all the cases listed in
table 1, where the legend of line styles and symbols is also given.

(a) (b) (c) 2.70

3.4 3.4
3.2 3.2
2.65
3.0 3.0
2.8 2.8 2.60
2.6 2.6
2.4 2.4 2.55
2.2 2.2
2.0 2.0 2.50
10 1 10 2 10 3 10 4 0 1000 2000 3000 4000 200 400 600 800 1000

F IGURE 3. (Colour online) Log-law indicator function β for (a) the highest-Reynolds-
number simulations, LM5200, LJ4200 and BPO4100, (b) simulations at Reτ = 5186, 2003
and 1000 (LM5200, HJ2000 and LM1000), and (c) simulation LM5200 along with the
expressions ((3.3), star) and ((3.4), diamond). In (a), the horizontal dashed line is at β =
1/κ = 1/0.384, and in (c) parameter values (see table 2) fitted from LM5200 are solid, and
those from Jiménez & Moser (2007), Mizuno & Jiménez (2011) are dashed. The linestyle
legend for (a) and (b) is given in table 1.

reported by Österlund et al. (2000), κ = 0.38, and Nagib & Chauhan (2008), κ =
0.384, in the zero-pressure-gradient boundary layer. However, the value of κ reported
here is smaller than κ = 0.40 measured by Bailey et al. (2014) in pipe flow. It should
be noted that the choice of the range for this curve fit is somewhat arbitrary, since
the indicator function is not exactly flat (figure 3c).
From an asymptotic analysis perspective, the log-law relation (3.1) is the lowest-
order truncation of a matched asymptotic expansion in 1/Reτ (Afzal & Yajnik 1973;
Jiménez & Moser 2007). Several higher-order representations of the mean velocity
in the overlap region have been evaluated based on experimental and DNS data in
boundary layers, channels and pipes (Buschmann & Gad-el-Hak 2003; Jiménez &
Moser 2007; Mizuno & Jiménez 2011). Here, we consider two such applications to
channels in the context of the LM5200 data.
Jiménez & Moser (2007) considered a higher-order truncation in which β has the
form +
1 y+
 
+ ∂U α1
β =y = + + α2 . (3.3)
∂y+ κ∞ Reτ Reτ
402 M. Lee and R. D. Moser

Equation (3.3) Equation (3.4)

Jiménez & Moser (2007) LM5200 Mizuno & Jiménez (2011) LM5200
κ∞ 0.402 Π 0.179 0.189
α1 150 a1 −12.4 −1.0
α2 1 0.2 a2 0.394 0.7
κ 0.3970 0.3867 κ 0.363 0.384
TABLE 2. Values of parameters in (3.3) and (3.4) appropriate for Reτ = 5186 as determined
by Jiménez & Moser (2007) and Mizuno & Jiménez (2011) and from the LM5200 data.

This formulation essentially allows for an Reτ dependence of κ = (1/κ∞ + α1 /Reτ )−1 ,
and introduces a linear dependence on y/δ = y+ /Reτ . Based on data from a simulation
at Reτ ≈ 1000 by Del Álamo et al. (2004, similar to LM1000) and from HJ2000
(see figure 3b), they determined the parameter values shown in table 2. Further, this
form and these values were found to be consistent with experimental measurements
by Christensen & Adrian (2001) at Reynolds numbers up to Reτ = 2433.
Mizuno & Jiménez (2011) considered a different higher-order asymptotic truncation,
for which β is given by

∂U + y+ y+2
β = y+ = + a2 2 . (3.4)
∂y+ κ(y − a1 )
+ Reτ

The second term was motivated by the form of a wake model that is quadratic
for small y/δ, where the coefficient is related to the wake parameter Π by
a2 = (12Π − 2)/κ. The term a1 is an offset (virtual origin) which accounts for
the presence of the viscous layer (Wosnik, Castillo & George 2000). To first order
in a1 /y+ , the offset is equivalent to including an additive a1 /κy+ term, which is
expected from the matched asymptotics. They fitted the inverse of the mean velocity
derivative to y+ /β from (3.4) using the experimental and DNS data mentioned
above and the experiments of Monty (2005), with up to Reτ = 3945, to determine a
Reynolds-number-dependent value of the parameters. When evaluated for Reτ = 5186,
these yield the parameter values shown in table 2.
These two higher-order truncations have also been fitted to the LM5200 data to
obtain values shown in table 2 that are significantly different from the previously
determined values. The expressions for β from (3.3) and (3.4) are plotted in figure 3(c)
with both sets of parameters. It is clear from this figure that the parameter values
obtained by Jiménez & Moser (2007) and Mizuno & Jiménez (2011) do not fit the
LM5200 data, but the parameters fitted to the LM5200 data in the log region match
the data equally well for both truncation forms. The reason for this disagreement with
the parameters from Jiménez & Moser (2007) is clear, since the Reynolds numbers
used in that study were not high enough to exhibit the logarithmic region observed
in LM5200, since y/δ = 0.16 is at y+ = 320 when Reτ = 2000. They appear to have
been fitting (3.3) to the outer-layer profile, and indeed in LM5200 there is a region
0.16 < y/δ < 0.45 in which β is approximately linear in y+ /Reτ , with a slope of one
(figure 3b), in agreement with α2 = 1.0.
In contrast, the data used by Mizuno & Jiménez (2011) included a channel Reynolds
number as high as Reτ = 3945, which should have exhibited a short nearly constant β
plateau as observed in LM5200. This would have been qualitatively different from the
 DNS of turbulent channel flow up to Reτ ≈ 5200 403
lower-Reynolds-number cases. This was not reported in that paper. However, at this
Reynolds number, the plotted values of κ and a1 (figure 5 in Mizuno & Jiménez 2011)
are significantly larger than the parameters for lower Reynolds number. Further, the
values for all three parameters obtained from the LM5200 data are within the indicated
uncertainties of the parameters from the Reτ = 3945 experimental data. Perhaps the
reported Reynolds number dependence of the parameters did not reflect a qualitative
change at the highest Reynolds number, because the fit was dominated by the more
numerous lower-Reynolds-number cases.
The apparent extent of the overlap region in the LM5200 case is not sufficient
to distinguish between the two asymptotic truncations, (3.3) and (3.4). Because a1
is so small, the primary distinction is in the lowest non-zero exponent on y+ /Reτ .
This may be of some importance because it determines the way in which the high-
Reynolds-number asymptote of constant β is approached. Unfortunately, this cannot
be determined using the available data.

3.2. Reynolds stress tensor

The non-zero components of the Reynolds stress tensor (the velocity component
variances and covariance) from the simulations and experiments are shown in figures
4–6. These figures show that there are some subtle inconsistencies among the three
highest-Reynolds-number simulations (solid lines in the figures) and the experimental
data.
First, the two cases LJ4200 and BPO4100 are nearly identical in Reynolds number,
but all three velocity variances are different between the two cases. The peak of
the streamwise variance (figure 4b) is approximately 1.4 % larger in LJ4200 than in
BPO4100. The peak varies with Reynolds number, as can be seen in the figure, but
this variation is logarithmic, and is too weak to explain the difference. Using the
simulations LM1000, HJ2000 and LM5200, the dependence of the peak in hu02 i+ on
Reτ was fitted to obtain

hu02 i+
max = 3.66 + 0.642 log(Reτ ), (3.5)

with R2 = 0.9995. This agrees well with the relationship hu02 i+

max = 3.63 + 0.65 log(Reτ )
suggested by Lozano-Durán & Jiménez (2014). The relationship (3.5) is plotted in
figure 4(c) along with the values of the actual peaks, including those for LJ4200 and
BPO4100. It is clear from this plot that the peak in BPO4100 is smaller than this
relationship implies for Reτ = 4100, while the peak from LJ4200 is consistent.
Also shown in figure 4(c) are experimental data for pipe flow by Hultmark et al.
(2010, 2012), which indicate that the inner-peak value of hu02 i+ does not continue to
increase with Re for Reτ > 5000. However, experimental data from boundary layers by
Hutchins et al. (2009) and Kulandaivelu (2011) do not show such a growth saturation.
The Reynolds numbers of the current simulations are unfortunately not high enough to
determine whether the growth of hu02 i+ will saturate in channel flows. Also remarkable
in figure 4(c) is how much lower the peak values are in the boundary layer data. This
is probably due to the resolution of the measurements (Hutchins et al. 2009).
In other y+ intervals (y+ < 8 and 25 < y+ < 200), the variance in the lower-
Reynolds-number case (BPO4100) is actually greater than the variance in the higher-
Reynolds-number flow (LJ4200). This appears to be inconsistent with the Reynolds
number trends among the other cases, for which hu02 i+ increases monotonically
with Reynolds number at constant y+ . Another inconsistency is apparent when the
streamwise velocity variance is examined as a function of y/δ (figure 4d, f ). Near the
404 M. Lee and R. D. Moser

(a) 10 (b) 10.0

9.5
8
9.0
6 8.5
8.0
4 7.5
7.0
2
6.5
0 6.0
10 0 10 1 10 2 10 3 10 4 10 1

(c) 10.0 (d) 10

9.5 LM1000
8
HJ1000
LJ4200
9.0 BPO4100
LM5200 6
MWT
8.5 PSP
4
8.0
7.5 2

7.0 0
10 3 10 4 10 5 10 –4 10 –3 10 –2 10 –1 10 0

(e) 8 ( f ) 2.0
6
4 1.5

2
1.0
0
–2 0.5
–4
–6 0
10 –4 10 –3 10 –2 10 –1 10 0 0.70 0.75 0.80 0.85 0.90 0.95 1.00

F IGURE 4. (Colour online) Variance of u: (a) as a function of y+ ; (b) zoom of (a) near the
peak; (c) dependence of maximum on Reτ ; solid line, relation (3.5); MWT, boundary layer
in the Melbourne Wind Tunnel (Hutchins et al. 2009; Kulandaivelu 2011); PSP, Princeton
Superpipe (Hultmark, Bailey & Smits 2010; Hultmark et al. 2012); (d) as a function of
y/δ; (e) test function for Townsend’s prediction; (f ) zoom of (d) near the centre of the
channel. The linestyle and symbol legend is given in table 1.

centreline (y/δ = 1), the variance from LJ4200 is significantly larger than from both
BPO4100 and LM5200, while the other simulations indicate that the variance should
be increasing slowly with Reynolds number. It appears that far from the wall, LJ4200
is affected by its relatively small domain size, as might be expected. Finally, the
experimental data with reported uncertainty, ±2 %, from SF4000 in figure 4(a) are
inconsistent with LJ4200 and BPO4100 in the region near the wall (y+ < 300 say).
The reason for this discrepancy is not clear, but it may be due to the difficulty of
measuring velocity fluctuations near the wall. Far from the wall, the experimental
 DNS of turbulent channel flow up to Reτ ≈ 5200 405

(a) 1.4 (b) 1.0

1.2
0.8
1.0
0.8 0.6
0.6 0.4
0.4
0.2
0.2
0 0
10 0 10 1 10 2 10 3 10 4 10 0 10 1 10 2 10 3 10 4

F IGURE 5. (Colour online) Variance of v (a) and covariance of u and v (b). The linestyle
and symbol legend is given in table 1.

data for these quantities are consistent with the simulations. Measurements are not
available at small enough y+ to compare peak values of hu02 i+ .
Similar inconsistencies are present among the high-Reynolds-number simulation
cases in the wall-normal and spanwise velocity variances. Around the peaks of both
hv 02 i+ and hw02 i+ , BPO4100 exceeds values from LJ4200, despite its somewhat lower
Reynolds number (figures 5a and 6a), while near the centre, these two cases are
in agreement. Only the Reynolds shear stress hu0 v 0 i+ is in agreement in these cases
across all y (figure 5b). Similarly to hu02 i+ , the experimental data from SF4000 for
hv 02 i+ and hu0 v 0 i+ with uncertainties of ±3 % and ±5 % respectively, are inconsistent
with both simulations in the region near the wall, where the experimental data appear
to be quite noisy. Possible reasons for the minor inconsistencies noted among the
DNS are discussed in § 3.3.
There are several anticipated high-Reynolds-number features to be examined in
the data. In particular, similarly to the log law, Townsend’s attached-eddy hypothesis
(Townsend 1976) implies that in the high-Reynolds-number limit, there is an interval
in y in which the Reynolds stress components satisfy
hu02 i+ = A1 − B1 log(y/δ), (3.6a)
hv 02 i+ = A2 , (3.6b)
hw02 i+ = A3 − B3 log(y/δ), (3.6c)
hu0 v 0 i+ = −1. (3.6d)
Consistent with these relations, both hv 02 i+ and −hu0 v 0 i+ are developing a flat region
as the Reynolds number increases (figure 5), though the maximum Reynolds shear
stress is well below one (0.96). Because the total stress
∂U
τtot = ν − hu0 v 0 i (3.7)
∂y
+
is known analytically (τtot = 1 − y/δ) from the mean momentum equation, and because
β varies little over a broad range of y (β = 2.6 ± 0.4 ≈ 1/κ for 30 < y+ < 0.75Reτ )
(see figure 3), the variation with Reynolds number of the Reynolds shear stress near
its maximum can be deduced easily:

+ y+ 1
τRS = −hu0 v 0 i+ ≈ 1 − − + for 30 < y+ < 0.75Reτ . (3.8)
Reτ κy
406 M. Lee and R. D. Moser

(a) 3.0 (b) 3.0

2.5 2.5
2.0 2.0
1.5 1.5
1.0 1.0
0.5 0.5
0 0
–0.5 –0.5
10 0 10 1 10 2 10 3 10 4 10 –4 10 –3 10 –2 10 –1 10 0

(c) 1.5 (d) 1.5

1.0 1.0

0.5 0.5

0 0

–0.5 –0.5

–1.0 –1.0

–1.5 –1.5
10 0 10 1 10 2 10 3 10 4 10 –4 10 –3 10 –2 10 –1 10 0

F IGURE 6. (Colour online) Variance of w (a) as a function of y+ , (b) as a function of

y/δ, and test function for Townsend’s prediction (c) as a function of y+ , (d) as a function
of y/δ. The linestyle and symbol legend is given in table 1.

From
√ this it is clear that the maximum Reynolds shear
√ stress is given by τRSmax ≈ 1 −
2/ κReτ , and that this maximum occurs at y+ ≈ Reτ /κ, as noted by Afzal (1982),
Morrison et al. (2004), Panton (2007) and Sillero et al. (2013). For the conditions
of LM5200 (Reτ = 5186, κ = 0.384), these estimates yield τRSmax ≈ 0.955 occurring
at y+ ≈ 116, in good agreement with the simulations. Further, the error in satisfying
τRS = 1 is less than  for a range of y+ that increases in size like Reτ for large Reτ .
Precisely,
s
Re Re 4
y+ −
τ τ
1 − τRS <  provided < 1− 2 . (3.9)
2 2  Reτ κ

Thus, for τRS to be within 5 % of one over a decade of variation of y+ would require
more than twice the Reynolds number of LM5200, and for it to be within 1 % at its
peak requires 20 times greater Reynolds number.
According to (3.6), both the variance of u and the variance of w would have a
logarithmic variation over some region of y. In figure 6, it appears that there is such a
logarithmic variation, even at Reynolds numbers as low as Reτ = 1000. The indicator
function y∂y hw02 i is approximately flat from y+ ≈ 100 to y+ ≈ 200 (figure 6c). The
corresponding curve fit is hw02 i+ = 1.08 − 0.387 log(y/δ), which is somewhat different
from the fit hw02 i+ = 0.8 − 0.45 log(y/δ) obtained by Sillero et al. (2013) in a boundary
 DNS of turbulent channel flow up to Reτ ≈ 5200 407

1.0

0.5

–0.5

–1.0
10 0 10 1 10 2 10 3 10 4

F IGURE 7. (Colour online) Balance of production and dissipation of turbulent kinetic

energy.

layer DNS. On the other hand, there is no apparent logarithmic region in the hu02 i+
profiles. The indicator function y∂y hu02 i+ (figure 4e) is not flat anywhere in the domain.
There is, however, a region (200 < y+ < 0.6Reτ ) where the dependence of the indicator
function on y is linear with relatively small slope, and the slope may be decreasing
with Reynolds number, though extremely slowly. In contrast, Hultmark et al. (2012,
2013) observed a logarithmic region in hu02 i+ over the range 800 < y+ < 0.15Reτ in
pipe flow with Reτ > 2 × 104 . The LM5200 simulation is not at high enough Reynolds
number to exhibit such a region if it occurs over the same range in y, since 0.15Reτ <
800 at Reτ = 5186.
As mentioned in § 2, the terms in the Reynolds stress transport equations are
not reported here, though the data are available at http://turbulence.ices.utexas.edu.
Of interest here, however, is the transport equation for the turbulent kinetic energy
K = hu0i u0i i/2. Hinze (1975) argues that at sufficiently high Reynolds number, there
is an intermediate region between inner and outer layers where the transport terms
in the kinetic energy equation are small compared with production, so that in this
region
PK ≈ , (3.10)
where PK is the production of kinetic energy and  is the dissipation. In the
formulation and analysis of turbulence models, (3.10) is often assumed to hold
in an overlap region between inner and outer layers (Durbin & Pettersson Reif
2010). The relative error in the balance of production and dissipation (PK / − 1)
for LM1000, HJ2000 and LM5200 is shown in figure 7. In all three cases, there
is a region y+ > 30 and y/δ < 0.6 in which the mismatch between production and
dissipation is of order 10 % or less, but there is no indication that the magnitude
of this mismatch is decreasing with Reynolds number. Indeed, there appears to be a
stable structure with a local minimum in PK / around y+ = 60 and a local maximum
around y+ = 300 followed by a gradual decline towards PK / = 1 with increasing y.
Presumably this decline will become more gradual with increasing Reynolds number
as y/δ = 0.6 increases in y+ .

3.3. Discrepancies between simulations

In §§ 3.1 and 3.2, it was shown that the LM5200 simulation has differences from the
two somewhat lower-Reynolds-number DNS LJ4200 and BPO4100, with discrepancies
among all three simulations larger than expected given the Reynolds number
408 M. Lee and R. D. Moser
differences and likely statistical errors. In the case of LJ4200, the reason for the
differences from LM5200 is most likely the small domain size of LJ4200 (four and
three times smaller in the streamwise and spanwise directions respectively). Indeed,
Lozano-Durán & Jiménez (2014) investigated the effects of the small domain size
at Reτ ≈ 950 by comparing with DNS in a domain consistent with the simulations
reported here. Consistent with observations in § 3.2, they observed discrepancies in
the velocity variances for y/δ < 0.25.
The differences between BPO4100 and both LJ4200 and LM5200 occur primarily in
the near-wall region, as shown in § 3.2. It seems likely that these differences are due to
numerical resolution limitations of BPO4100. As indicated in table 1, the grid spacing
in BPO4100 is approximately the same as LM5200 in the spanwise direction and
approximately 25 % finer in the streamwise direction. However, in BPO4100 a second-
order finite difference approximation is used rather than the spectral method used in
the simulations reported here. The effective resolution of such low-order schemes is
significantly less at the same grid spacing. One way to see this is to consider the
error incurred when differentiating a sine function of different wavenumbers, which is
given by (k) = 1 − sin(k∆)/k∆. If one insists on limiting error to no more than 10 %
(for example), then the wavenumber needs to be limited to one fourth of the highest
wavenumber that can be represented on the grid. Thus, at this level of error, the finite
difference approximation of the derivative has four times coarser effective resolution
than a spectral method on the same grid.
Of course, there is much more to solving the Navier–Stokes equations than
representing the derivative, and an appropriate error limit for the derivative is not
clear. Nonetheless, in DNS, a common rule of thumb is that second-order finite
difference has between two and four times coarser effective resolution than a Fourier
spectral method on the same grid. In a study of the effects of resolution in DNS of
low-Reynolds-number (Reτ = 180) channel flow, Oliver et al. (2014) found that with
the Fourier/B-spline numerical representation used here, coarsening the resolution
by a factor of two relative to the nominal resolution (nominal is comparable in
wall units to that used for LM5200) results in changes of several per cent in the
velocity variances near the wall. Further, Vreman & Kuerten (2014) investigated
differences between DNS using a Fourier–Chebyshev spectral method and DNS
using a fourth-order staggered finite difference method, which is higher order and
higher resolution than the method used for BPO4100. They reported a 1 % lower
peak root-mean-square (r.m.s.) u0 for the fourth-order method. These results indicate
that the lower effective resolution in BPO4100 relative to LM5200 is a plausible
cause of the minor inconsistencies between these two simulations. To determine this
definitively it would be necessary to redo the BPO4100 simulation with twice the
resolution in each direction, or more, which is out of the scope of the current study.

3.4. Energy spectral density

One of the properties of high-Reynolds-number wall-bounded turbulence is the
separation of scales between the near-wall and outer-layer turbulence. For the LM5200
case, this separation of scales can be seen in the one-dimensional velocity spectra.
For example, the premultiplied spectral energy density of the streamwise velocity
fluctuations is shown as a function of y in figure 8. The premultiplied spectrum
kE(k, y+ ) is the energy density per log k, and so, in the logarithmic wavenumber
scale used in figure 8, it indicates the scales at which the energy resides. The energy
spectral density of u in the streamwise direction has two distinct peaks, one at
kx δ = 40, y+ = 13 and one at kx δ = 1, y+ = 400. To the best of our knowledge,
 DNS of turbulent channel flow up to Reτ ≈ 5200 409

(a) (b)
10 3 10 3

10 2 10 2

10 1 10 1

10 0 10 0
10 0 10 1 10 2 10 3 10 0 10 1 10 2 10 3

(c) (d)
10 3 10 3

10 2 10 2

10 1 10 1

10 0 10 0
10 0 10 1 10 2 10 3 10 0 10 1 10 2 10 3

F IGURE 8. (Colour online) Wavenumber-premultiplied energy spectral density from

LM5200: (a) kx Euu /u2τ , (b) kz Euu /u2τ , (c) −kx Euv /u2τ , (d) −kz Euv /u2τ .

such distinct peaks have only previously been observed in high-Reynolds-number

experimental data (Hutchins & Marusic 2007; Monty et al. 2009; Marusic, Mathis &
Hutchins 2010a; Marusic et al. 2010b). An even more vivid double peak is visible in
the spanwise spectrum of u (figure 8b), with peaks at kz δ = 250, y+ = 13 and kz δ = 6,
y+ = 1000.
Similarly, there are weak double peaks in the cospectrum of uv, as shown in
figure 8(c,d). In the LM5200 case, distinct inner and outer peaks were not observed
in the spectral density of v 2 and w2 (not shown). It thus appears that it is primarily
the streamwise velocity fluctuations that are exhibiting inner/outer scale separation at
this Reynolds number.
As mentioned earlier, the scaling analysis of Perry et al. (1986) suggests that
at high Reynolds number the energy spectral density of the streamwise velocity
fluctuations varies as kx−1 in the overlap region where both inner and outer scaling
are valid. However, until now, the kx−1 region has been elusive in simulations,
presumably because the Reynolds numbers have not been high enough. It is only in
high-Reynolds-number experiments (Nickels et al. 2005, 2007; Rosenberg et al. 2013)
that a kx−1 region has previously been observed in the streamwise velocity spectrum.
In the LM5200 simulation, the premultiplied energy spectral density does indeed
exhibit a plateau in the region 90 6 y+ 6 170 and 6 6 kx δ 6 10 (figure 9a). Further,
the magnitude of the premultiplied spectrum at the plateau agrees with the value
observed experimentally, approximately 0.8 in wall units, in boundary layers (Nickels
et al. 2005) and in pipe flow (Rosenberg et al. 2013). However, the plateau in the
410 M. Lee and R. D. Moser

(a) 1.0 (b) 4.0

3.5
3.0
2.5
0.8
2.0
1.5
1.0
0.6 0.5
0
10 0 10 1 10 2 10 3

0.4 (c) 2.0

1.5

0.2 1.0

0.5
0
10 0 10 1 10 2 0
10 0 10 1 10 2 10 3

F IGURE 9. (Colour online) The k−1 region: (a) kx Euu /u2τ at y+ = 90–170, (b) kz Euu /u2τ at
y+ = 3–14, (c) kz Euu /u2τ at y+ = 111–141.

experimental premultiplied spectrum in the pipe flow of Rosenberg et al. (2013) is

only observed for Reτ 6 3300. From the current results, we cannot determine whether
the occurrence or extent of the plateau region might change at higher Reynolds
number in channels.
A similar scaling analysis also suggests that there should be a plateau in the
premultiplied spanwise spectrum, and indeed an even broader plateau appears for
5 6 kz δ 6 30, as observed previously by Hoyas & Jiménez (2008) and Sillero et al.
(2013). Interestingly, the plateau also occurs in the viscous sublayer region (figure 9b).
In the viscous sublayer the streamwise velocity fluctuations are dominated by the
well-known streaky structures, with a characteristic spacing of 1z+ ≈ 100. This
is evidenced by the large high-wavenumber peaks in the spanwise spectra around
kz /δ ≈ 300 (kz+ ≈ 2π/1z+ ). At much larger scales (much lower wavenumbers),
the viscous layer is driven by the larger-scale turbulence further from the wall. The
plateau in the spanwise premultiplied spectrum in the viscous layer is thus a reflection
of the spectral plateau further from the wall (figure 9c).
Another experimentally observed feature of the streamwise premultiplied spectrum
is the presence of two local maxima, with one peak occurring on either side of the
plateau (Guala, Hommema & Adrian 2006; Kunkel & Marusic 2006). However, this
bimodal feature was called into question when it was noted that the low-wavenumber
peak, which occurs at kx δ of order one, is at low enough wavenumber to be affected
by the use of Taylor’s hypothesis to infer the spatial spectrum from the measured
temporal spectrum (Del Álamo & Jiménez 2009; Moin 2009). Indeed, in the HJ2000
simulation, it was found that the streamwise spatial spectrum did not display a low-
wavenumber peak, while a spectrum determined from a time series did (Del Álamo
& Jiménez 2009). In the LM5200 case, this bimodal feature does occur, as can be
 DNS of turbulent channel flow up to Reτ ≈ 5200 411
seen in figure 9(a). However, it is not clear whether this is a general high-Reynolds-
number feature, or whether it is characteristic of intermediate Reynolds numbers. From
figure 8, it is clear that the high- and low-wavenumber peaks in kx Euu (kx , y+ ) at around
y+ = 100 are simply the upper and lower edges of the near wall and outer peaks in
kx Euu (kx , y+ ) respectively. As the Reynolds number increases, the inner peak will move
to larger kx and the outer peak to larger y+ . As the outer peak moves to larger y+ , it
seems likely that the inner and outer peaks will not overlap at all in y, resulting in no
bimodal premultiplied spectrum at any y. For example, the peaks are better separated
in the premultiplied spanwise spectrum, and there are no bimodal kz premultiplied
spectra (figure 9b).

4. Discussion and conclusion

The DNS of turbulent channel flow at Reτ = 5186 that is reported here has
been shown to be a reliable source of data on high-Reynolds-number wall-bounded
turbulence, and a wealth of statistical data from this flow is available online at
http://turbulence.ices.utexas.edu. In particular, the resolution is consistent with
or better than accepted standards for wall-bounded turbulence DNS, statistical
uncertainties are generally small (of order 1 % or less), statistical data are consistent
with Reynolds-number trends in DNS at Reτ 6 2000, and with experimental data,
within reasonable tolerance, and, finally, the flow is statistically stationary to high
accuracy. Further, as recapped below, the simulation exhibits many characteristics
of high-Reynolds-number wall-bounded turbulence, making this simulation a good
resource for the study of such high-Reynolds-number flows. Several conclusions can
thus be drawn about high-Reynolds-number turbulent channel flow, as discussed below.
At high Reynolds number, it is expected that a region of logarithmic variation
will exist in the mean velocity. The current simulation exhibits an unambiguous
logarithmic region with von Kármán constant κ = 0.384 ± 0.004. This is the first
such unambiguous simulated logarithmic profile that the authors are aware of. At
high Reynolds number, it is also expected (Townsend 1976) that the variance of the
velocity fluctuations will have a region of logarithmic variation. This was found to
be true for the spanwise fluctuations, but not the streamwise fluctuations. Indeed, the
streamwise variance shows no sign of converging towards a logarithmic variation
with increasing Reynolds number. Finally, it is often assumed that at high Reynolds
number the production of turbulent kinetic energy will locally balance its dissipation
over some overlap region between inner and outer layers. However, this was observed
only approximately in the current simulation, with 10 % accuracy and with no sign
that this mismatch is declining with Reynolds number. This balance of production
and dissipation thus appears to be just an imperfect approximation.
One of the most important characteristics of high-Reynolds-number wall-bounded
turbulence is a distinct separation in scale between turbulence near the wall and
far from the wall. This is observed in the streamwise and spanwise premultiplied
spectral density of the streamwise velocity fluctuations and in the cospectrum of
streamwise and wall-normal fluctuations. Moreover, a short k−1 region is observed
in the streamwise spectra and there is a wider region in the spanwise spectra, as
predicted from scaling analysis (Perry et al. 1986), and as observed experimentally.
Finally, the bimodal structure in the premultiplied spectrum that has been observed
experimentally (Guala et al. 2006; Kunkel & Marusic 2006) has also been observed
here, despite suggestions that the experimental observations were an artefact of the
use of Taylor’s hypothesis, which is not used here. However, it seems likely that this
two-peak structure will not continue as the Reynolds number increases and the inner
and outer peaks in the premultiplied spectra become further separated in k and y.
412 M. Lee and R. D. Moser
Acknowledgements
The work presented here was supported by the National Science Foundation under
award number (OCI-0749223) and the Argonne Leadership Computing Facility at
Argonne National Laboratory under Early Science Program (ESP) and Innovative and
Novel Computational Impact on Theory and Experiment Program (INCITE) 2013.
We wish to thank N. Malaya and R. Ulerich for software engineering assistance. We
are also thankful to M. P. Schultz and K. A. Flack for providing their experimental
data.

REFERENCES

A FZAL , N. 1976 Millikan’s argument at moderately large Reynolds number. Phys. Fluids 19 (4),
600–602.
A FZAL , N. 1982 Fully developed turbulent flow in a pipe: an intermediate layer. Ing.-Arch. 52,
355–377.
A FZAL , N. & YAJNIK , K. 1973 Analysis of turbulent pipe and channel flows at moderately large
Reynolds number. J. Fluid Mech. 61, 23–31.
BAILEY, S. C. C., VALLIKIVI , M., H ULTMARK , M. & S MITS , A. J. 2014 Estimating the value of
von Kármán’s constant in turbulent pipe flow. J. Fluid Mech. 749, 79–98.
B ERNARDINI , M., P IROZZOLI , S. & O RLANDI , P. 2014 Velocity statistics in turbulent channel flow
up to Reτ = 4000. J. Fluid Mech. 742, 171–191.
B ORRELL , G., S ILLERO , J. A. & J IMÉNEZ , J. 2013 A code for direct numerical simulation of
turbulent boundary layers at high Reynolds numbers in BG/P supercomputers. Comput. Fluids
80, 37–43.
B OTELLA , O. & S HARIFF , K. 2003 B-spline methods in fluid dynamics. Intl J. Comput. Fluid Dyn.
17 (2), 133–149.
B USCHMANN , M. H. & G AD - EL -H AK , M. 2003 Generalized logarithmic law and its consequences.
AIAA J. 41 (1), 40–48.
C HRISTENSEN , K. T. & A DRIAN , R. J. 2001 Statistical evidence of hairpin vortex packets in wall
turbulence. J. Fluid Mech. 431, 433–443.
C OMTE -B ELLOT, G. 1963 Contribution à l’étude de la turbulence de conduite. PhD thesis, University
of Grenoble, France.
D EAN , R. B. & B RADSHAW, P. 1976 Measurements of interacting turbulent shear layers in a duct.
J. Fluid Mech. 78, 641–676.
D E G RAAFF , D. B. & E ATON , J. K. 2000 Reynolds-number scaling of the flat-plate turbulent boundary
layer. J. Fluid Mech. 422, 319–346.
D EL Á LAMO , J. C. & J IMÉNEZ , J. 2009 Estimation of turbulent convection velocities and corrections
to Taylor’s approximation. J. Fluid Mech. 640, 5–26.
D EL Á LAMO , J. C., J IMÉNEZ , J., Z ANDONADE , P. & M OSER , R. D. 2004 Scaling of the energy
spectra of turbulent channels. J. Fluid Mech. 500, 135–144.
D IXIT, S. A. & R AMESH , O. N. 2013 On the k1−1 scaling in sink-flow turbulent boundary layers.
J. Fluid Mech. 737, 329–348.
D URBIN , P. A. & P ETTERSSON R EIF , B. A. 2010 Statistical Theory and Modeling for Turbulent
Flows. Wiley.
E L K HOURY, G. K., S CHLATTER , P., N OORANI , A., F ISCHER , P. F., B RETHOUWER , G. &
J OHANSSON , A. V. 2013 Direct numerical simulation of turbulent pipe flow at moderately
high Reynolds numbers. Flow Turbul. Combust. 91 (3), 475–495.
F ERNHOLZ , H. H. & F INLEY, P. J. 1996 The incompressible zero-pressure-gradient turbulent boundary
layer: an assessment of the data. Prog. Aerosp. Sci. 32 (8), 245–311.
G UALA , M., H OMMEMA , S. E. & A DRIAN , R. J. 2006 Large-scale and very-large-scale motions in
turbulent pipe flow. J. Fluid Mech. 554, 521–542.
H INZE , J. O. 1975 Turbulence. McGraw-Hill.
 DNS of turbulent channel flow up to Reτ ≈ 5200 413
H OYAS , S. & J IMÉNEZ , J. 2006 Scaling of the velocity fluctuations in turbulent channels up to
Reτ = 2003. Phys. Fluids 18 (1), 011702.
H OYAS , S. & J IMÉNEZ , J. 2008 Reynolds number effects on the Reynolds-stress budgets in turbulent
channels. Phys. Fluids 20 (10), 101511.
H ULTMARK , M., BAILEY, S. C. C. & S MITS , A. J. 2010 Scaling of near-wall turbulence in pipe
flow. J. Fluid Mech. 649, 103–113.
H ULTMARK , M., VALLIKIVI , M., BAILEY, S. C. C. & S MITS , A. J. 2012 Turbulent pipe flow at
extreme Reynolds numbers. Phys. Rev. Lett. 108, 094501.
H ULTMARK , M., VALLIKIVI , M., BAILEY, S. C. C. & S MITS , A. J. 2013 Logarithmic scaling of
turbulence in smooth- and rough-wall pipe flow. J. Fluid Mech. 728, 376–395.
H UTCHINS , N., C HAUHAN , K., M ARUSIC , I., M ONTY, J. & K LEWICKI , J. 2012 Towards reconciling
the large-scale structure of turbulent boundary layers in the atmosphere and laboratory.
Boundary-Layer Meteorol. 145 (2), 273–306.
H UTCHINS , N. & M ARUSIC , I. 2007 Large-scale influences in near-wall turbulence. Phil. Trans. R.
Soc. Lond. A 365 (1852), 647–664.
H UTCHINS , N., N ICKELS , T. B., M ARUSIC , I. & C HONG , M. S. 2009 Hot-wire spatial resolution
issues in wall-bounded turbulence. J. Fluid Mech. 635, 103–136.
J IMÉNEZ , J. & M OSER , R. D. 2007 What are we learning from simulating wall turbulence? Phil.
Trans. R. Soc. Lond. A 365 (1852), 715–732.
J OHANSSON , A. V. & A LFREDSSON , P. H. 1982 On the structure of turbulent channel flow. J. Fluid
Mech. 122, 295–314.
J OHNSON , R. W. 2005 Higher order B-spline collocation at the Greville abscissae. Appl. Numer.
Maths 52 (1), 63–75.
K IM , J., M OIN , P. & M OSER , R. 1987 Turbulence statistics in fully developed channel flow at low
Reynolds number. J. Fluid Mech. 177, 133–166.
K IM , K. C. & A DRIAN , R. J. 1999 Very large-scale motion in the outer layer. Phys. Fluids 11 (2),
417.
K ULANDAIVELU , V. 2011 Evolution and structure of zero pressure gradient turbulent boundary layer.
PhD thesis, University of Melbourne.
K UNKEL , G. J. & M ARUSIC , I. 2006 Study of the near-wall-turbulent region of the high-Reynolds-
number boundary layer using an atmospheric flow. J. Fluid Mech. 548, 375–402.
K WOK , W. Y., M OSER , R. D. & J IMÉNEZ , J. 2001 A critical evaluation of the resolution properties
of B-spline and compact finite difference methods. J. Comput. Phys. 174 (2), 510–551.
L EE , M., M ALAYA , N. & M OSER , R. D. 2013 Petascale direct numerical simulation of turbulent
channel flow on up to 786K cores. In Proceedings of SC13: International Conference for
High Performance Computing, Networking, Storage and Analysis. ACM.
L EE , M., U LERICH , R., M ALAYA , N. & M OSER , R. D. 2014 Experiences from leadership computing
in simulations of turbulent fluid flows. Comput. Sci. Engng 16 (5), 24–31.
L OZANO -D URÁN , A. & J IMÉNEZ , J. 2014 Effect of the computational domain on direct simulations
of turbulent channels up to Reτ = 4200. Phys. Fluids 26 (1), 011702.
M ARUSIC , I., M ATHIS , R. & H UTCHINS , N. 2010a High Reynolds number effects in wall turbulence.
Intl J. Heat Fluid Flow 31 (3), 418–428.
M ARUSIC , I., M C K EON , B. J., M ONKEWITZ , P. A., NAGIB , H. M., S MITS , A. J. & S REENIVASAN ,
K. R. 2010b Wall-bounded turbulent flows at high Reynolds numbers: recent advances and
key issues. Phys. Fluids 22 (6), 065103.
M ARUSIC , I., M ONTY, J. P., H ULTMARK , M. & S MITS , A. J. 2013 On the logarithmic region in
wall turbulence. J. Fluid Mech. 716, R3.
M ILLIKAN , C. B. 1938 A critical discussion of turbulent flows in channels and circular tubes. In
Proceedings of the 5th International Congress for Applied Mechanics, pp. 386–392. Wiley.
M IZUNO , Y. & J IMÉNEZ , J. 2011 Mean velocity and length-scales in the overlap region of wall-
bounded turbulent flows. Phys. Fluids 23 (8), 085112.
M OIN , P. 2009 Revisiting Taylor’s hypothesis. J. Fluid Mech. 640, 1–4.
414 M. Lee and R. D. Moser
M ONTY, J. P. 2005 Developments in smooth wall turbulent duct flows. PhD thesis, University of
Melbourne.
M ONTY, J. P. & C HONG , M. S. 2009 Turbulent channel flow: comparison of streamwise velocity
data from experiments and direct numerical simulation. J. Fluid Mech. 633, 461–474.
M ONTY, J. P., H UTCHINS , N., N G , H. C. H., M ARUSIC , I. & C HONG , M. S. 2009 A comparison
of turbulent pipe, channel and boundary layer flows. J. Fluid Mech. 632, 431–442.
M ORRISON , J. F., M C K EON , B. J., J IANG , W. & S MITS , A. J. 2004 Scaling of the streamwise
velocity component in turbulent pipe flow. J. Fluid Mech. 508, 99–131.
M OSER , R. D., K IM , J. & M ANSOUR , N. N. 1999 Direct numerical simulation of turbulent channel
flow up to Reτ = 590. Phys. Fluids 11 (4), 943–945.
N AGIB , H., C HRISTOPHOROU , C., R EUDI , J.-D., M ONKEWITZ , P., Ö STERLUN , J. & G RAVANTE ,
S. 2004 Can we ever rely on results from wall-bounded turbulent flows without direct
measurements of wall shear stress. In 24th AIAA Aerodynamic Measurement Technology and
Ground Testing Conference, Portland, Oregon, p. 2392. American Institute of Aeronautics and
Astronautics (AIAA).
NAGIB , H. M. & C HAUHAN , K. A. 2008 Variations of von Kármán coefficient in canonical flows.
Phys. Fluids 20 (10), 101518.
N ICKELS , T., M ARUSIC , I., H AFEZ , S. & C HONG , M. S. 2005 Evidence of the k1−1 law in a
high-Reynolds-number turbulent boundary layer. Phys. Rev. Lett. 95 (7), 074501.
N ICKELS , T. B., M ARUSIC , I., H AFEZ , S., H UTCHINS , N. & C HONG , M. S. 2007 Some predictions
of the attached eddy model for a high Reynolds number boundary layer. Phil. Trans. R. Soc.
Lond. A 365 (1852), 807–822.
O LIVER , T. A., M ALAYA , N., U LERICH , R. & M OSER , R. D. 2014 Estimating uncertainties in
statistics computed from direct numerical simulation. Phys. Fluids 26 (3), 035101.
Ö STERLUND , J. M., J OHANSSON , A. V., N AGIB , H. M. & H ITES , M. H. 2000 A note on the
overlap region in turbulent boundary layers. Phys. Fluids 12 (1), 1–4.
PANTON , R. L. 2007 Composite asymptotic expansions and scaling wall turbulence. Phil. Trans. R.
Soc. Lond. A 365 (1852), 733–754.
P ERRY, A. E., H ENBEST, S. & C HONG , M. S. 1986 A theoretical and experimental study of wall
turbulence. J. Fluid Mech. 165, 163–199.
ROSENBERG , B. J., H ULTMARK , M., VALLIKIVI , M., B AILEY, S. C. C. & S MITS , A. J. 2013
Turbulence spectra in smooth- and rough-wall pipe flow at extreme Reynolds numbers.
J. Fluid Mech. 731, 46–63.
S CHULTZ , M. P. & F LACK , K. A. 2013 Reynolds-number scaling of turbulent channel flow. Phys.
Fluids 25 (2), 025104.
S ILLERO , J. A., J IMÉNEZ , J. & M OSER , R. D. 2013 One-point statistics for turbulent wall-bounded
flows at Reynolds numbers up to δ + ≈ 2000. Phys. Fluids 25, 105102.
S MITS , A. J. & M ARUSIC , I. 2013 Wall-bounded turbulence. Phys. Today 66 (9), 25–30.
S MITS , A. J., M C K EON , B. J. & M ARUSIC , I. 2011 High-Reynolds number wall turbulence. Annu.
Rev. Fluid Mech. 43 (1), 353–375.
S PALART, P. R. & A LLMARAS , S. R. 1992 A one-equation turbulence model for aerodynamic flows.
In 30th Aerospace Sciences Meeting and Exhibit, p. 439. American Institute of Aeronautics
and Astronautics.
S PALART, P. R., M OSER , R. D. & ROGERS , M. M. 1991 Spectral methods for the Navier–Stokes
equations with one infinite and two periodic directions. J. Comput. Phys. 96 (2), 297–324.
T OWNSEND , A. A. 1976 The Structure of Turbulent Shear Flow, 2nd edn. Cambridge University
Press.
V REMAN , A. W. & K UERTEN , J. G. M. 2014 Comparison of direct numerical simulation databases
of turbulent channel flow at Reτ = 180. Phys. Fluids 26 (1), 015102.
W EI , T. & W ILLMARTH , W. W. 1989 Reynolds-number effects on the structure of a turbulent channel
flow. J. Fluid Mech. 204, 57–95.
W ESTERWEEL , J., E LSINGA , G. E. & A DRIAN , R. J. 2013 Particle image velocimetry for complex
and turbulent flows. Annu. Rev. Fluid Mech. 45 (1), 409–436.
 DNS of turbulent channel flow up to Reτ ≈ 5200 415
W INKEL , E. S., C UTBIRTH , J. M., C ECCIO , S. L., P ERLIN , M. & D OWLING , D. R. 2012 Turbulence
profiles from a smooth flat-plate turbulent boundary layer at high Reynolds number. Exp. Therm.
Fluid Sci. 40, 140–149.
W OSNIK , M., C ASTILLO , L. & G EORGE , W. K. 2000 A theory for turbulent pipe and channel flows.
J. Fluid Mech. 421, 115–145.
W U , X., BALTZER , J. R. & A DRIAN , R. J. 2012 Direct numerical simulation of a 30R long turbulent
pipe flow at R+ = 685: large- and very large-scale motions. J. Fluid Mech. 698, 235–281.
Z ANOUN , E.-S., D URST, F. & NAGIB , H. 2003 Evaluating the law of the wall in two-dimensional
fully developed turbulent channel flows. Phys. Fluids 15 (10), 3079.
Z ANOUN , E.-S., N AGIB , H. & D URST, F. 2009 Refined cf relation for turbulent channels and
consequences for high-Re experiments. Fluid Dyn. Res. 41 (2), 021405.