824 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 5, NO.

6, DECEMBER 1997

Hashed and Hierarchical Timing Wheels: Efficient

Data Structures for Implementing a Timer Facility
George Varghese and Anthony Lauck

Abstract— The performance of timer algorithms is crucial The performance of algorithms to implement a timer module
to many network protocol implementations that use timers for becomes an issue when any of the following are true.
failure recovery and rate control. Conventional algorithms to
implement an Operating System timer module take O (n ) time to • The algorithm is implemented by a processor that is
start or maintain a timer, where n is the number of outstanding interrupted each time a hardware clock ticks, and the
timers: this is expensive for large n. This paper shows that by interrupt overhead is substantial.
using a circular buffer or timing wheel, it takes O (1) time to • Fine granularity timers are required.
start, stop, and maintain timers within the range of the wheel.
• The average number of outstanding timers is large.
Two extensions for larger values of the interval are described. In
the first, the timer interval is hashed into a slot on the timing If the hardware clock interrupts the host every tick, and the
wheel. In the second, a hierarchy of timing wheels with different interval between ticks is in the order of microseconds, then the
granularities is used to span a greater range of intervals. The interrupt overhead is substantial. Most host operating systems
performance of these two schemes and various implementation
tradeoffs are discussed. We have used one of our schemes to
offer timers of coarse (milliseconds or seconds) granularity.
replace the current BSD UNIX callout and timer facilities. Our Alternately, in some systems finer granularity timers reside
new implementation can support thousands of outstanding timers in special-purpose hardware. In either case, the performance
without much overhead. Our timer schemes have also been of the timer algorithms will be an issue as they determine the
implemented in other operating systems and network protocol latency incurred in starting or stopping a timer and the number
packages.
of timers that can be simultaneously outstanding.
Index Terms— Callout facilities, hashed wheels, hierarchical As an example, consider communications between members
wheels, protocol implementations, Timers, Timer Facilities. of a distributed system. Since messages can be lost in the
underlying network, timers are needed at some level to trigger
I. INTRODUCTION retransmissions. A host in a distributed system can have
several timers outstanding. Consider, for example, a server
I N a centralized or distributed system, we need timers for
the following.
Failure Recovery: Several kinds of failures cannot be de-
with 200 connections and 3 timers per connection. Further, as
networks scale to gigabit speeds, both the required resolution
tected asynchronously. Some can be detected by periodic and the rate at which timers are started and stopped will
checking (e.g., memory corruption) and such timers always increase. Several recent network implementations (e.g., [6])
expire. Other failures can only be inferred by the lack of have been tuned to send packets at a rate of 25 000–40 000
some positive action (e.g., message acknowledgment) within a packets per second.
specified period. If failures are infrequent, these timers rarely Some network implementations (e.g., the BSD TCP im-
expire. plementation) do not use a timer per packet; instead, only
Algorithms in Which the Notion of Time or Relative Time is a few timers are used for the entire networking package.
Integral: Examples include algorithms that control the rate The BSD TCP implementation gets away with two timers
of production of some entity (process control, rate-based because the TCP implementation maintains its own timers for
flow control in communications), scheduling algorithms, and all outstanding packets, and uses a single kernel timer as a
algorithms to control packet lifetimes in computer networks. clock to run its own timers. TCP maintains its packet timers in
These timers almost always expire. the simplest fashion: whenever its single kernel timer expires,
it ticks away at all its outstanding packet timers. For example,
Manuscript received November 1, 1996; revised June 19, 1997; approved many TCP implementations use two timers: a 200-ms timer
by IEEE/ACM TRANSACTIONS ON NETWORKING Editor L. Peterson. An earlier and a 500-ms timer.
version of this paper appeared in Proc. 11th ACM Symp. on Operating Systems
Principles, Nov. 1987. The work of G. Varghese was supported by the Office The naive method works reasonably well if the granularity
of Naval research under an ONR Young Investigator Award and by the of timers is low and losses are rare. However, it is desirable
National Science Foundation under Grant NCR 940997. to improve the resolution of the retransmission timer to allow
G. Varghese was with Digital Equipment Corporation, Littleton, MA USA.
He is now with the Department of Computer Science, Washington University, speedier recovery. For example, the University of Arizona has
St. Louis, MO 63130 USA (e-mail: [email protected]). a new TCP implementation called TCP Vegas [4] that performs
A. Lauck was with Digital Equipment Corporation, Littleton, MA better than the commonly used TCP Reno. One of the reasons
USA. He is now at P.O. Box 59, Warren, VT 05674 USA (e-mail:
[email protected]). TCP Reno has bad performance when experiencing losses is
Publisher Item Identifier S 1063-6692(97)07265-8. the coarse granularity of the timeouts.
1063–6692/97$10.00  1997 IEEE
VARGHESE AND LAUCK: HASHED AND HIERARCHICAL TIMING WHEELS 825

Besides faster error recovery, fine granularity timers also TABLE I

allow network protocols to more accurately measure small AN EXAMPLE OF THE PARAMETERS OF THE TIMER MODULE THAT
A NETWORKING APPLICATION WOULD CONSIDER IMPORTANT
intervals of time. For example, accurate estimates of round trip
delay are important for the TCP congestion control algorithm Routine Critical Parameter
STARTTIMER Latency (average and worst-case)
[14] and the Scalable Reliable Multicast (SRM) framework STOPTIMER Latency (average and worst-case)
[11] that is implemented in the Wb conferencing tool [16]. PERTICKBOOKKEEPING Latency (average)
Finally, many multimedia applications routinely use timers, EXPIRYPROCESSING None
and the number of such applications is increasing. An example
can be found in Siemens’ CHANNELS run time system for TABLE II
multimedia [3] where each audio stream uses a timer with LATENCY METRICS FOR THREE PREVIOUSLY USED SCHEMES. NOTE THAT
granularity that lies between 10 and 20 ms. For multime- O O n
STOPTIMER IS (1) FOR UNBALANCED TREES AND [log ( )] FOR BALANCED
TREES; BALANCED TREE IMPLEMENTATIONS HAVE THE SLOWEST STOPTIMER
dia and other real-time applications, it is important to have BECAUSE OF THE NEED TO REBALANCE THE TREE AFTER A DELETION
worst-case bounds on the processing time to start and stop
Scheme STARTTIMER STOPTIMER PERTICK
timers.
1 O(1) O(1) O(n)
Besides networking applications, process control and other 2 O(n) O(1) O(1)
real-time applications will also benefit from large numbers 3 O(log (n)) O(log (n)) or O(1) O(1)
of fine granularity timers. Also, the number of users on a
system may grow large enough to lead to a large number of
outstanding timers. This is the reason cited (for redesigning caller of the routine blocks until the routine completes.
the timer facility) by the developers of the IBM VM/XA SP1 Both the average and worst case latency are of interest.
operating system [10]. For example, a client application that implements a trans-
In the following sections, we will describe a family of port protocol may find that space is cheap and the critical
schemes for efficient timer implementations based on a data parameters for each routine in the timer module are as shown
structure called a timing wheel. We will also describe perfor- in Table I.
mance results based on a UNIX implementation, and survey The performance measures important for the client applica-
some of the systems that have implemented timer packages tions should be used to choose among timer algorithms.
based on the ideas in this paper.
III. EXISTING TIMER SCHEMES
II. MODEL There are two standard schemes.
Our model of a timer module has the following four
component routines. A. Scheme 1—Straightforward
STARTTIMER (Interval, RequestId, ExpiryAction): The client Here [22] STARTTIMER finds a memory location and sets
calls this routine to start a timer that will expire after “Interval” that location to the specified timer interval. Every units,
units of time. The client supplies a RequestId which is used PERTICKBOOKKEEPING will decrement each outstanding timer;
to distinguish this timer from other timers that the client has if any timer becomes zero, EXPIRYPROCESSING is called.
outstanding. Finally, the client can specify what action must be This scheme is extremely fast except for per tick bookkeep-
taken on expiry: for instance, calling a client-specified routine, ing. It also uses one record per outstanding timer, the minimum
or setting an event flag. space possible. Its performance is summarized in Table II. It
STOPTIMER (RequestId): This routine uses its knowledge of is appropriate if:
the client and RequestId to locate the timer and stop it.
• there are only a few outstanding timers;
PERTICKBOOKKEEPING: Let the granularity of the timer be
• most timers are stopped within a few ticks of the clock;
units. Then every units this routine checks whether any
• PERTICKBOOKKEEPING is done with suitable performance
outstanding timers have expired; if this is the case, it calls
by special-purpose hardware.
STOPTIMER, which in turn calls the next routine.
EXPIRYPROCESSING: This routine does the ExpiryAction Note that instead of doing a Decrement, we can store the
specified in the STARTTIMER call. absolute time at which timers expire and do a Compare. This
The first two routines are activated on client calls while option is valid for all timer schemes we describe; the choice
the last two are invoked on timer ticks. The timer is often an between them will depend on the size of the time-of-day field,
external hardware clock. the cost of each instruction, and the hardware on the machine
The following two performance measures can be used to implementing these algorithms. In this paper we will use the
choose between the various algorithms described in the rest of Decrement option, except when describing Scheme 2.
this paper. Both of them are parameterized by , the average
(or worst-case) number of outstanding timers. B. Scheme 2—Ordered List
1) Space: The memory required for the data structures used Here [22] PERTICKBOOKKEEPING latency is reduced at the
by the timer module. expense of STARTTIMER performance. Timers are stored in an
2) Latency: The time between the invoking of a routine in ordered list. Unlike Scheme 1, we will store the absolute time
the timer module and its completion, assuming that the at which the timer expires, and not the interval before expiry.
826 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 5, NO. 6, DECEMBER 1997

Fig. 1. Timer queue example used to illustrate Scheme 2. Fig. 3. Analogy between a timer module and a sorting module.

STOPTIMER need not search the list if the list is doubly

linked. When STARTTIMER inserts a timer into the ordered
list, it can store a pointer to the element. STOPTIMER can then
use this pointer to delete the element in (1) time from the
Fig. 2. A G/G/Inf/Inf queuing model of a timer module. Note that s(t) is doubly linked list. This can be used by any timer scheme.
the density function of interval between starting and stopping (or expiration) If Scheme 2 is implemented by a host processor, the
of a timer.
interrupt overhead on every tick can be avoided if there is
hardware support to maintain a single timer. The hardware
The timer that is due to expire at the earliest time is stored at timer is set to expire at the time at which the timer at the head
the head of the list. Subsequent timers are stored in increasing of the list is due to expire. The hardware intercepts all clock
order as shown in Fig. 1. ticks and interrupts the host only when a timer actually expires.
In Fig. 1, the lowest timer is due to expire at absolute time Unfortunately, some processor architectures do not offer this
10 h, 23 min, and 12 s. capability. Algorithms similar to Scheme 2 are used by both
Because the list is sorted, PERTICKBOOKKEEPING need only VMS and UNIX in implementing their timer modules. The
increment the current time of day, and compare it with the performance of the two schemes is summarized in Table II.
head of the list. If they are equal, or the time of day is greater, As for Space, Scheme 1 needs the minimum space possible;
it deletes that list element and calls EXPIRYPROCESSING. It Scheme 2 needs extra space for the forward and back
continues to delete elements at the head of the list until the pointers between queue elements.
expiry time of the head of the list is strictly less than the time
of day.
STARTTIMER searches the list to find the position to insert IV. SORTING TECHNIQUES AND
the new timer. In the example, STARTTIMER will insert a new TIME-FLOW MECHANISMS
timer due to expire at 10:24:01 between the second and third
elements. A. Sorting Algorithms and Priority Queues
The worst-case latency to start a timer is O(n). The average Scheme 2 reduced PERTICKBOOKKEEPING latency at the
latency depends on the distribution of timer intervals (from expense of STARTTIMER by keeping the timer list sorted.
time started to time stopped), and the distribution of the arrival Consider the relationship between timer and sorting algorithms
process according to which calls to STARTTIMER are made. depicted in Fig. 3. However, consider the following.
Interestingly, this can be modeled (Fig. 2) as a single queue
• In a typical sort, all elements are input to the module when
with infinite servers; this is valid because every timer in the
the sort begins; the sort ends by outputting all elements in
queue is essentially decremented (or served) every timer tick.
sorted order. A timer module performs a more dynamic
It is shown in [17], that we can use Little’s result to obtain
sort because elements arrive at different times and are
the average number in the queue; also the distribution of the
output at different times.
remaining time of elements in the timer queue seen by a
• In a timer module, the elements to be “sorted” change
new request is the residual life density of the timer interval
their value over time if we store the interval. This is not
distribution.
true if we store the absolute time of expiry.
If the arrival distribution is Poisson, the list is searched from
the head, and reads and writes both cost one unit, then the A data structure that allows “dynamic” sorting is a priority
average cost of insertion for negative exponential and uniform queue [7]. A priority queue allows elements to be inserted
timer interval distributions is shown in [17] to be and deleted; it also allows the smallest element in the set to
be found. A timer module can use a priority queue, and do
— negative exponential PERTICKBOOKKEEPING only on the smallest timer element.
— uniform. 1) Scheme 3—Tree-Based Algorithms: A linked list
(Scheme 2) is one way of implementing a priority queue. For
Results for other timer interval distributions can be com- large , tree-based data structures are better. These include
puted using a result in [17]. For a negative exponential unbalanced binary trees, heaps, post-order and end-order trees,
distribution we can reduce the average cost to by and leftist-trees [7], [26]. They attempt to reduce the latency
searching the list from the rear. In fact, if timers are always in Scheme 2 for STARTTIMER from to .
inserted at the rear of the list, this search strategy yields an In [18] it is reported that this difference is significant for
(1) STARTTIMER latency. This happens, for instance, if all large , and that unbalanced binary trees are less expensive
timers intervals have the same value. However, for a general than balanced binary trees. Unfortunately, unbalanced binary
distribution of the timer interval, we assume the average trees easily degenerate into a linear list; this can happen, for
latency of insertion is . instance, if a set of equal timer intervals are inserted.
VARGHESE AND LAUCK: HASHED AND HIERARCHICAL TIMING WHEELS 827

We will lump these algorithms together as Scheme 3:

tree-based algorithms. The performance of Scheme 3 is sum-
marized in Table II.

B. Discrete Event Simulation

In discrete event simulations [19], all state changes in the
system take place at discrete points in time. An important part
of such simulations are the event-handling routines or time-
flow mechanisms. When an event occurs in a simulation, it
may schedule future events. These events are inserted into
some list of outstanding events. The simulation proceeds by Fig. 4. Timing wheel mechanism used in logic simulation [21].
processing the earliest event, which in turn may schedule
further events. The simulation continues until the event list is current cycle are removed from the overflow list and inserted
empty or some condition (e.g., clock MaxSimulationTime) into the array of lists. This is implemented in TEGAS-2
holds. [21].
There are two ways to find the earliest event and update The array can be conceptually thought of as a timing
the clock. wheel; every time we step through locations, we rotate
1) The earliest event is immediately retrieved from some the wheel by incrementing the number of cycles. A problem
data structure (e.g., a priority queue [7]) and the clock with this implementation is that as time increases within
jumps to the time of this event. This is embodied in a cycle and we travel down the array, it becomes more
simulation languages like GPSS [12] and SIMULA [9]. likely that event records will be inserted in the overflow list.
2) In the simulation of digital circuits, it is often sufficient Other implementations [15] reduce (but do not completely
to consider event scheduling at time instants that are avoid) this effect by rotating the wheel half-way through the
multiples of the clock interval, say . Then, after the array.
program processes an event, it increments the clock In summary, we note that time flow algorithms used for
variable by until it finds any outstanding events at digital simulation can be used to implement timer algorithms;
the current time. It then executes the event(s). This conversely, timer algorithms can be used to implement time
is embodied in languages for digital simulation like flow mechanisms in simulations. However, there are differ-
TEGAS [21] and DECSIM [15]. ences to note.
We have already seen that algorithms used to implement • In digital simulations, most events happen within a short
the first method are applicable for timer algorithms: these interval beyond the current time. Since timing wheel
include linked lists and tree-based structures. What is more implementations rarely place event notices in the overflow
interesting is that algorithms for the second method are also list, they do not optimize this case. This is not true for a
applicable. Translated in terms of timers, the second method general-purpose timer facility.
for PERTICKBOOKKEEPING is: “Increment the clock by the • Most simulations ensure that if two events are scheduled
clock tick. If any timer has expired, call EXPIRYPROCESSING.” to occur at the same time, they are removed in FIFO
An efficient and widely used method to implement the sec- order. Timer modules need not meet this restriction.
ond method is the so-called timing-wheel [21], [24] technique. • Stepping through empty buckets on the wheel represents
In this method, the data structure into which timers are inserted overhead for a digital simulation. In a timer module,
is an array of lists, with a single overflow list for timers beyond we have to increment the clock anyway on every tick.
the range of the array. Consequently, stepping through empty buckets on a clock
In Fig. 4, time is divided into cycles; each cycle is units tick does not represent significant extra overhead if it is
of time. Let the current number of cycles be . If the current done by the same entity that maintains the current time.
time pointer points to element , the current time is . • Simulation languages assume that canceling event notices
The event notice corresponding to an event scheduled to arrive is very rare. If this is so, it is sufficient to mark the notice
within the current cycle (e.g., at time , for integer as “canceled” and wait until the event is scheduled; at that
between 0 and ) is inserted into the list pointed to by point, the scheduler discards the event. In a timer module,
the th element of the array. Any event occurring beyond STOPTIMER may be called frequently; such an approach
the current cycle is inserted into the overflow list. Within a can cause the memory needs to grow unboundedly beyond
cycle, the simulation increments the current time until it finds the number of timers outstanding at any time.
a nonempty list; it then removes and processes all events in the We will use the timing-wheel method below as a point of
list. If these schedule future events within the current cycle, departure to describe further timer algorithms.
such events are inserted into the array of lists; if not, the new
events are inserted into the overflow list.
The current time pointer is incremented modulo . When V. SCHEME 4—BASIC SCHEME
it wraps to 0, the number of cycles is incremented, and the We describe a simple modification of the timing-wheel
overflow list is checked; any elements due to occur in the algorithm. If we can guarantee that all timers are set for
828 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 5, NO. 6, DECEMBER 1997

Fig. 5. Array of lists used by Scheme 4 for timer intervals up to MaxInterval. Fig. 6. Array of lists used by Schemes 5 and 6 for arbitrary-sized timers:
basically a hash table.

periods less than MaxInterval, this modified algorithm takes

latency for STARTTIMER, STOPTIMER, and also for VI. EXTENSIONS
PERTICKBOOKKEEPING. Let the granularity of the timer be 1
unit. The current time is represented in Fig. 5 by a pointer to an A. Extension 1—Hashing
element in a circular buffer with dimensions [0, MaxInterval The previous scheme has an obvious analogy to inserting
1]. an element in an array using the element value as an index. If
To set a timer at units past current time, we index (Fig. 5) there is insufficient memory, we can hash the element value
into Element (mod MaxInterval), and put the timer at the to yield an index.
head of a list of timers that will expire at a time = Current Time For example, if the table size is a power of 2, an arbitrary
units. Each tick we increment the current timer pointer size timer can easily be divided by the table size; the remainder
(mod MaxInterval) and check the array element being pointed (low order bits) is added to the current time pointer to yield
to. If the element is 0 (no list of timers waiting to expire), the index within the array. The result of the division (high
no more work is done on that timer tick. But if it is nonzero, order bits) is stored in a list pointed to by the index.
we do expiry processing on all timers that are stored in that In Fig. 6, let the table size be 256 and the timer be a 32-
list. Thus, the latency for STARTTIMER is . The cost of bit timer. The remainder on division is the last 8 bits. Let
PERTICKBOOKKEEPING is except when timers expire, but the value of the last 8 bits be 20. Then the timer index is 10
this is the best possible. If the timer lists are doubly linked, (Current Time Pointer) 20 (remainder) 30. The 24 high
and, as before, we store a pointer to each timer record, then order bits are then inserted into a list that is pointed to by the
the latency of STOPTIMER is also . 30th element.
This is basically a timing-wheel scheme where the wheel Other methods of hashing are possible. For example, any
turns one array element every timer unit, as opposed to rotating function that maps a timer value to an array index could be
every MaxInterval or MaxInterval/2 units [21]. This guarantees used. We will defend our choice at the end of this subsection.
that all timers within MaxInterval of the current time will Next, there are two ways to maintain each list.
be inserted in the array of lists; this is not guaranteed by 1) Scheme 5—Hash Table With Sorted Lists: Here each
conventional timing wheel algorithms [15], [21]. list is maintained as a ordered list exactly as in Scheme
2. STARTTIMER can be slow because the 24 bit quantity must
In sorting terms, this is similar to a bucket sort [7] that
be inserted into the correct place in the list. Although the
trades off memory for processing. However, since the timers
worst-case latency for STARTTIMER is still , the average
change value every time instant, intervals are entered as offsets
latency can be . This is true if TableSize, and
from the current time pointer. It is sufficient if the current time
if the hash function (which is TimerValue mod TableSize)
pointer increases every time instant.
distributes timer values uniformly across the table. If so, the
A bucket sort sorts elements in time using
average size of the list that the th element is inserted into
buckets, since all buckets have to be examined. This is is TableSize [7]. Since TableSize, the average
inefficient for large . In timer algorithms, however, latency of STARTTIMER is . How well this hash actually
the crucial observation is that some entity needs to do distributes depends on the arrival distribution of timers to this
work per tick to update the current time; it costs only a few module, and the distribution of timer intervals.
more instructions for the same entity to step through an empty PERTICKBOOKKEEPING must increment the current time
bucket. What matters, unlike the sort, is not the total amount pointer. If the value stored in the array element being pointed
of work to sort elements, but the average (and worst-case) to is zero, there is no more work. Otherwise, as in Scheme 2,
part of the work that needs to be done per timer tick. the top of the list is decremented. If the timer at the top of
Still memory is finite: it is difficult to justify 2 words the list expires, EXPIRYPROCESSING is called and the top list
of memory to implement 32 bit timers. One solution is element is deleted. Once again, PERTICKBOOKKEEPING takes
to implement timers within some range using this scheme average and worst-case latency except when multiple
and the allowed memory. Timers greater than this value are timers are due to expire at the same instant, which is the best
implemented using, say, Scheme 2. Alternately, this scheme we can do.
can be extended in two ways to allow larger values of the Finally, if each list is doubly linked and STARTTIMER stores
timer interval with modest amounts of memory. a pointer to each timer element, STOPTIMER takes time.
VARGHESE AND LAUCK: HASHED AND HIERARCHICAL TIMING WHEELS 829

A pleasing observation is that the scheme reduces to Scheme

2 if the array size is 1. In terms of sorting, Scheme 5 is similar
to doing a bucket sort on the low order bits, followed by an
insertion sort [7] on the lists pointed to by each bucket.
2) Scheme 6—Hash Table with Unsorted Lists: If a worst-
case STARTTIMER latency of is unacceptable, we can
maintain each time list as an unordered list instead of an
ordered list. Thus, STARTTIMER has a worst case and average
latency of . But the per-tick bookkeeping now takes
longer. Every timer tick, we increment the pointer (mod
TableSize); if there is a list there, we must decrement the high
order bits for every element in the array, exactly as in Scheme
1. However, if the hash table has the property described above,
then the average size of the list will be . Fig. 7. Hierarchical set of arrays of lists used by Scheme 7 to “map” time
We can make a stronger statement about the average behav- more efficiently.
ior regardless of how the hash distributes. Notice that every
TableSize ticks we decrement once all timers that are still As an example, consider Fig. 7. Let the current time be 11
living. Thus, for timers, we do TableSize work on average days, 10 h, 24 min, 30 s. Then to set a timer of 50 min and 45
per tick. If TableSize then we do work on average
s, we first calculate the absolute time at which the timer will
per tick. If all timers hash into the same bucket, then every
expire. This is 11 days, 11 h, 15 min, 15 s. Then we insert
TableSize ticks we do work, but for intermediate ticks
the timer into a list beginning 1 (11–10 hrs) element ahead of
we do work.
the current hour pointer in the hour array. We also store the
Thus, the hash distribution in Scheme 6 only controls
remainder (15 min and 15 s) in this location. We show this in
the variance of the latency of PERTICKBOOKKEEPING, and
Fig. 7, ignoring the day array which does not change during
not the average latency. Since the worst-case latency of
the example.
PERTICKBOOKKEEPING is always (all timers expire at
The seconds array works as usual: every time the hardware
the same time), we believe that the choice of hash function for
clock ticks, we increment the second pointer. If the list pointed
Scheme 6 is insignificant. Obtaining the remainder after divid-
ing by a power of 2 is cheap, and consequently recommended. to by the element is nonempty, we process all elements in the
Further, using an arbitrary hash function to map a timer value list using EXPIRYPROCESSING. However, the other three arrays
into an array index would require PERTICKBOOKKEEPING to work slightly differently.
compute the hash on each timer tick, which would make it Even if there are no timers requested by the user of the
more expensive. service, there will always be a 60-s timer that is used to update
We discuss implementation strategies for Scheme 6 in the minute array, a 60-min timer to update the hour array, and
Appendix A. a 24-h timer to update the day array. For instance, every time
the 60-s timer expires, we will increment the current minute
timer, do any required expiry processing for the minute timers,
and reinsert another 60-s timer.
B. Extension 2—Exploiting Hierarchy Returning to the example, if the timer is not stopped,
The last extension of the basic scheme exploits the concept eventually the hour timer will reach 11. When the hour timer
of hierarchy. To represent the number 1 000 000 we need only reaches 11, the list is examined. The expiry processing routine
seven digits instead of 1 000 000 because we represent num- will insert the remainder of the seconds (15) in the minute
bers hierarchically in units of 1’s, 10’s, 100’s etc. Similarly, array, 15 elements after the current minute pointer (0). Of
to represent all possible timer values within a 32-bit range, we course, if the minutes remaining were zero, we could go
do not need a 2 element array. Instead we can use a number directly to the second array. At this point, the table will look
of arrays, each of different granularity. For instance, we can like Fig. 8.
use four arrays as follows: Eventually, the minute array will reach the 15th element; as
• a 100-element array in which each element represents a part of EXPIRYPROCESSING, we will move the timer into the
day; second array 15 s after the current value. Fifteen seconds later,
• a 24-element array in which each element represents an the timer will actually expire, at which point the user-specified
hour; EXPIRYPROCESSING is performed.
• a 60-element array in which each element represents a What are the performance parameters of this scheme?
minute; STARTTIMER: Depending on the algorithm, we may need
• a 60-element array in which each element represents a time, where is the number of arrays in the hierarchy, to
second. find the right table to insert the timer and to find the remaining
Thus, instead of million locations time. A small number of levels should be sufficient to cover
to store timers up to 100 days, we need only 100 24 60 the timer range with an allowable amount of memory; thus
60 244 locations. should be small (say, ).
830 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 5, NO. 6, DECEMBER 1997

modes: one for hour timers, one for minute timers, etc. This
reduces PERTICKBOOKKEEPING overhead further at the cost
of a loss in precision of up to 50% (e.g., a 1-min and 30-s
timer that is rounded to 1 min). Alternately, we can improve
the precision by allowing just one migration between adjacent
lists.
Scheme 7 has an obvious analogy to a radix sort [7]. We
discuss implementation strategies for Scheme 7 in Appendix
A.

VII. UNIX IMPLEMENTATION

A. Costello of Washington University has implemented
Fig. 8. The previous example, after the hour component of the timer expires
(using Scheme 7). [8] a new version of the BSD UNIX callout and timer
facilities. Current BSD kernels take time proportional to the
number of outstanding timers to set or cancel timers. The new
STOPTIMER: Once again, this can be done in time if implementation, which is based on Scheme 6, takes constant
all lists are doubly linked. time to start, stop, and maintain timers; this leads to a highly
PERTICKBOOKKEEPING: It is useful to compare this to the scalable design that can support thousands of outstanding
corresponding value in Scheme 6. Both have the same average timers without much overhead.
latency of for sufficiently large array sizes but the In the existing BSD implementation, each callout is repre-
constants of complexity are different. More precisely: let sented by a callout structure containing a pointer to the
be the average timer interval (from start to stop or expiry); let function to be called (c func), a pointer to the function’s
be the total amount of array elements available; and let argument (c arg), and a time (c time) expressed in units
be the total number of levels in the hierarchy. of clock ticks. Outstanding callouts are kept in a linked list,
The total work done in Scheme 6 for such an average sized sorted by their expiration times. The c time member of each
timer is callout structure is differential, not absolute—the first callout
in the list stores the number of ticks from now until expiration,
and each subsequent callout in the list stores the number of
where is a constant denoting the cost of decrementing the ticks between its own expiration and the expiration of its
high order bits, indexing, etc., in Scheme 6. If a timer lives predecessor.
for units of time, it will be decremented times. In BSD UNIX, Callouts are set and canceled using routines
And in Scheme 7 it is bounded from above by called timeout() and untimeout(), respectively.
The routine timeout(func, arg, time) registers
func(arg) to be called at the specified time; untime-
where represents the cost of finding the next list to migrate out(func, arg) cancels the callout with matching
to, and the cost of migration, in Scheme 7; is the maximum function and argument. Because the calltodo list must
number of lists to migrate between. be searched linearly, both operations take time proportional to
The average cost per unit time for an average of timers the number of outstanding callouts. Interrupts are locked out
then becomes for the duration of the search.
— Scheme The Costello implementation is based on Scheme 6 de-
scribed above. Unfortunately, the existing timeout() inter-
— Scheme
face in BSD does not allow the passing of handles, which was
The choice between Scheme 6 and Scheme 7 will depend used in all our schemes to quickly cancel a timer. The Costello
on the parameters above. Since and will not be implementation used two solutions to this problem. For calls
drastically different, for small values of and large values using the existing interface, a search for a callout given a
of , Scheme 6 can be better than Scheme 7 for both function pointer and argument is done using a hash table. A
STARTTIMER and PERTICKBOOKKEEPING. However, for large second solution was also implemented: a new interface func-
values of and small values of , Scheme 7 will have a tion was defined for removing a callout (unsetcallout())
better average cost (latency) for PERTICKBOOKKEEPING but a that takes a handle as its only argument. This allows existing
greater cost for STARTTIMER latency. code to use the old interface and new applications to use the
W. Nichols has pointed out that if the timer precision is new interface. The performance difference between these two
allowed to decrease with increasing levels in the hierarchy, approaches appears to be slight, so the hash table approach
then we need not migrate timers between levels. For instance, appears to be preferable.
in the example above, we would round off to the nearest hour In the new implementation, the timer routines are guaranteed
and only set the timer in hours. When the hour timer goes off, to lock out interrupts only for a small, bounded amount of time.
we do the user-specified EXPIRYPROCESSING without migrating The new implementation also extends the setitimer()
to the minute array. Essentially, we now have different timer interface to allow a process to have multiple outstanding
VARGHESE AND LAUCK: HASHED AND HIERARCHICAL TIMING WHEELS 831

Fig. 9. Real-time performance comparison of BSD UNIX callout implementations. Note that the new callout implementations using timing wheels take
constant time. By contrast, the traditional BSD implementation takes time that increases linearly with the number of outstanding callouts.

timers, thereby reducing the need for users to maintain their VIII. LATER WORK
own timer packages. The changes to the BSD kernel are small A preliminary version of the work described in this paper
(548 lines of code added, 80 removed) and are available on was first described in [25]. Since then, a number of systems
the World Wide Web. The details of this new implementation have built timer implementations based on this approach, and
are described elsewhere [8]; the written report contains several there have been a few extensions of the basic approach.
important implementation details that are not described here. 1) Systems that Use Timing Wheels: Some well-known net-
work protocol implementations have used the timing wheel
A. Performance ideas described in this paper. These include the fast TCP
The performance of Scheme 6 was tested (using the Costello implementation in [6] and the X-kernel timer facility [1]. The
implementation). The tests took advantage of the new inter- efficient user level protocol implementation in [23] mentions
face extensions that allow a single process to have multiple the possible use of timing wheels but did not do an implemen-
outstanding callouts. We quote the following results from [8]. tation. We also know of commercial networking products that
Three kernels were tested on a Sun 4/360. The first kernel use timing wheels as part of their operating system. These
used the timeout() interface to the old callout facility. The include DEC’s Gigaswitch [20] and Siemens’ CHANNELS
second kernel used the existing interface but used the new run time system [2].
callout facility (and a hash table). The last kernel used the 2) Timing Wheel Extensions: Brown [5] extended the idea
new setcallout() interface (which allows handles) to the of hashed timing wheels to what he calls calendar queues.1 The
new callout facility. major difference is that calendar queue implementations also
In each test, one process created a number of outstanding periodically resize the wheel in order to reduce the overhead2
timers set for random times far in the future, causing a number of stepping through empty buckets. For timer applications, the
of outstanding callouts. It then created one more timer, and clock time must be incremented on every clock tick anyway;
repeatedly set it for a random time farther in the future thus adding a few instructions to step through empty buckets is
than the others, causing repeated calls to untimeout() and not significant. Davison [10] describes a timer implementation
timeout() (or unsetcallout() and setcallout(), for the IBM VM/XA SP1 operating system based on calendar
depending on which kernel was being used). The results queues. The empirical improvement in per tick bookkeeping
(Fig. 9) show that the time for the original callout facility (due to resizing the wheel periodically) does not appear to
increases linearly with the number of outstanding callouts, warrant the extra complexity of resizing.
whereas the time for the replacement callout facility is constant
with respect to the number of outstanding callouts, for both
the old interface (using hashing) and the new interface (using
1 Some authors refer to timing wheels as a variant of calendar queues; given
handles). The new interface performs very slightly better,
the dates of invention and publication, it is perhaps more accurate to say that
and provides guaranteed constant time operations, but the calendar queues are an extension of timing wheels.
old interface is needed for compatibility with the rest of the 2 The improvement is not worst-case and is only demonstrated empirically
kernel. for certain benchmarks.
832 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 5, NO. 6, DECEMBER 1997

IX. AN ALGORITHMIC VIEW to standard priority queue implementations like heaps [7].
From an algorithmic point of view, a timing wheel is just However, the constants appear to be better for Scheme 7.
a priority queue [7]. It appears to be just an application
of bucket sorting techniques to priority queues. However,
bucket sorting cannot be used efficiently for all priority queue
X. SUMMARY AND CONCLUSIONS
implementations. Timing wheels work efficiently only for
priority queue applications that satisfy the following bounded In this paper, we have examined the relationship between
monotonicity property: any elements inserted into the priority sorting algorithms, time flow mechanisms in discrete event
queue are within Max of the last minimum extracted. simulations, and timer algorithms. We have extended the
If this condition is satisfied and the inserted values are all timing wheel mechanism used in logic simulation to yield 3
integers, then we can implement the priority queue using a timer algorithms (Schemes 5–7) that have constant complexity
circular array of size Max. New elements are inserted into the for setting, stopping, and maintaining a timer. The extensions
circular array based on the difference between their value and include rotating the timing wheel every clock tick, having
the current minimum element. A pointer is kept to the last separate overflow lists per bucket, and using a hierarchical
minimum extracted. To find the new minimum at any point, set of timing wheels (Scheme 7): the extensions are necessary
we simply advance the pointer till an array location is found because the requirements of a scheduler in a logic simulation
that contains a valid element. This is exactly what is done in and those of a general timer module are different.
In choosing between schemes, we believe that Scheme 1 is
Scheme 4, where Max corresponds to MaxInterval.
appropriate in some cases because of its simplicity, limited use
It is easy to see what goes wrong if the monotonicity
of memory, and speed in starting and stopping timers. Scheme
condition is not satisfied. If we can insert an element that
2 is useful in a host that has hardware to maintain the clock
is smaller than the last minimum extracted, then we cannot
and a single timer. Although it takes time to start a timer,
advance the pointer to find the new minimum value; the pointer
the host is not interrupted every clock tick.
may have to backtrack, leading to a potential search of the
In a host without hardware support for timers, we believe
entire array.
Schemes 2 and 3 are inappropriate because of the cost of
Even with the monotonicity condition, the wheel approach
STARTTIMER when there are a large number of outstanding
to priority queues still requires stepping through empty buck-
timers. Clearly, this is not uncommon in hosts that have a
ets. However, the nice thing about timer applications is that
significant amount of real-time activity or have several open
many systems must maintain the time of day anyway, and
communication links.
thus the cost of stepping through empty buckets is amortized Scheme 4 is useful when most timers are within a small
over the existing cost of incrementing the time-of-day clock. range of the current time. For example, it could be used
This example illustrates how an algorithm, that may have a by a networking module that is maintaining its own timers.
poor algorithmic complexity when considered in isolation, can Scheme 5 depends too much on the hash distribution (for a
be very efficient when considered as part of a system, where fast STARTTIMER) to be generally useful. However, a variant
parts of the algorithms cost can be charged to other system of this scheme has been implemented in the X-kernel [1].
components. For a general timer module, similar to the operating system
The bounded monotonicity condition is satisfied by other facilities found in UNIX or VMS, that is expected to work
algorithmic applications. For example, for graphs with integer well in a variety of environments, we recommend Scheme 6
edge weights, the Dijkstra algorithm for shortest paths and or 7. The UNIX results described in this paper are encouraging,
Prim’s algorithm for minimum spanning trees [7] both satisfy and show that it is possible to support thousands of outstanding
the monotonicity condition with Max equal to the maximum timers at low overhead using Scheme 6.
edge weight. While it has been observed before [7] that these If the amount of memory required for an efficient implemen-
two algorithms can benefit from bucket sorting using a linear tation of Scheme 6 is a problem, Scheme 7 can be pressed into
array, the required size of the linear array was supposed to be service. Scheme 7, however, will need a few more instructions
the equal to the cost of the largest shortest cost path between in STARTTIMER to find the correct table to insert the timer.
any two nodes. Our observation shows that a circular array Both Schemes 6 and 7 can be completely or partially (see
of size equal to the maximum edge weight suffices. While Appendix A) implemented in hardware using some auxiliary
this is a mild observation, it does reduce the memory needs memory to store the data structures. If a host had such
of networking implementations that use Dijkstra’s algorithm hardware support, the host software would need time
and integer edge weights [13]. To the best of our knowledge, to start and stop a timer and would not need to be interrupted
the bounded monotonicity condition has not been described every clock tick.
before in the literature. Finally, we note that designers and implementers have
The efficiency of the hashed wheel solution (Scheme 6) assumed that protocols that use a large number of timers are
for larger timer values is based on bounding the number of expensive and perform poorly. This is an artifact of existing
timers and doing an amortized analysis. This does not appear implementations and operating system facilities. Given that a
to have any direct correspondence with bucket sorting. The large number of timers can be implemented efficiently, we
hierarchical scheme (Scheme 7) uses essentially logarithmic hope this will no longer be an issue in the design of protocols
time to insert an element; thus it is comparable in complexity for distributed systems.
VARGHESE AND LAUCK: HASHED AND HIERARCHICAL TIMING WHEELS 833

APPENDIX A ACKNOWLEDGMENT
HARDWARE ASSIST B. Spinney suggested extending Scheme 4 to Scheme 5.
Since the cost of handling clock interrupts becomes more H. Wilkinson independently thought of exploiting hierarchy
significant for fine granularity (e.g., microseconds) timers, it in maintaining timer lists. J. Forecast helped the authors
may be necessary to employ special-purpose hardware assist. implement an early version of Scheme 6. A. Black commented
In the extreme, we can use a timer chip which maintains all the on an earlier version and helped improve the presentation. A.
data structures (say in Scheme 6) and interrupts host software Black, B. Spinney, H. Wilkinson, S. Glaser, W. Nichols, P.
only when a timer expires. Koning, A. Kirby, M. Kempf, and C. Kaufman (all at DEC)
Another possibility is a chip (actually just a counter) that were a pleasure to discuss these schemes with. The authors are
steps through the timer arrays, and interrupts the host only if grateful to E. Cooper, M. Bjöorkman, C. Thekath, V. Seidel,
there is work to be done. When the host inserts a timer into an B. Souza, and A. Costello for giving information about their
empty queue pointed to by array element , it tells the chip implementations.
about this new queue. The chip then marks as “busy.” As
before, the chip scans through the timer arrays every clock tick.
During its scan, when the chip encounters a “busy” location, it
REFERENCES
interrupts the host and gives the host the address of the queue
that needs to be worked on. Similarly when the host deletes [1] M. Björkman, personal communication.
a timer entry from some queue and leaves behind an empty [2] S. Boecking and V. Seidel, “TIP’s protocol run-time system,” in
EFOCN’94, June 1994.
queue it needs to inform the chip that the corresponding array [3] S. Boecking, V. Seidel, and P. Vindeby, “CHANNELS—A run-time
location is no longer “busy.” system for multimedia protocols,” in ICCCN’95, Sept. 1995.
[4] L. Brakmo, S. O Malley, and L. Peterson, “TCP Vegas: New techniques
Note that the synchronization overhead is minimal because for congestion detection and avoidance,” in Proc. ACM SIGCOMM’94,
the host can keep the actual timer queues in its memory which London, England.
the chip need not access, and the chip can keep the timing O
[5] R. Brown, “Calendar queues: A fast (1) priority queue implementation
for the simulation event set problem,” Commun. ACM, vol. 31, no. 10,
arrays in its memory, which the host need not access. The pp. 1220–1227, Oct. 1988.
only communication between the host and chip is through [6] D. D. Clark, V. Jacobson, J. Romkey, and H. Salwen, “An analysis of
interrupts. TCP processing overhead,” IEEE Commun. Mag., vol. 27, no. 6, pp.
23–29, June 1989.
In Scheme 6, the host is interrupted an average of [7] T. Cormen, C. Leiserson, and R. Rivest, Introduction to Algorithms.
times per timer interval, where is the average timer interval Cambridge, MA: MIT Press/McGraw-Hill, 1990.
[8] A. Costello and G. Varghese, “Redesigning the BSD callout and timeout
and is the number of array elements. In Scheme 7, the facilities,” Dept. Computer Science, Washington Univ., St. Louis, MO,
host is interrupted at most times, where is the number Tech. Rep. 95-23, Sept. 1995.
of levels in the hierarchy. If and are small and is [9] O.-J. Dahl, B. Myhrhaug, and K. Nygaard, SIMULA’67 Common Base
Language, Norwegian Computing Center, Forksningveien, 1B, Oslo 3,
large, the interrupt overhead for such an implementation can Pub. S22.
be made negligible. [10] G. Davison, “Calendar p’s and q’s,” Commun. ACM, vol. 32, no. 10,
Finally, we note that conventional hardware timer chips use pp. 1241–1242, Oct. 1989.
[11] S. Floyd, V. Jacobson, S. McCanne, C. Liu, and L. Zhang, “A reliable
Scheme 1 to maintain a small number of timers. However, if multicast framework for light-weight sessions and application level
Schemes 6 and 7 are implemented as a single chip that operates framing,” in Proc ACM SIGCOMM’95, Boston, MA.
on a separate memory (that contains the data structures), [12] General Purpose Simulation System 360—User’s Manual, IBM Corp.,
White Plains, NY, Pub. H20-0326, 1968.
then there is no a priori limit on the number of timers [13] International Organization for Standardization (ISO), “Protocol for pro-
that can be handled by the chip. Clearly the array sizes viding the connectionless-model network service,” Draft International
Standard 8473, Mar. 1988.
need to be parameters that must be supplied to the chip on [14] V. Jacobson, “ Congestion avoidance and control,” in Proc. ACM
initialization. SIGCOMM’88, Stanford, CA.
[15] M. A. Kearney, “DECSIM: A multi-level simulation system for digital
design,” in 1984 Int. Conf. Computer Design.
[16] S. Mccanne, “A distributed whiteboard for network conferencing,”
University of California–Berkeley, Computer Networks Term project
APPENDIX B CS 268, May 1992.
SYMMETRIC MULTIPROCESSING [17] C. M. Reeves, “Complexity analysis of event set algorithms,” Computer
J., vol. 27, no. 1, 1984.
If the host consists of a set of processors, each of which [18] B. Myhrhaug, Sequencing Set Efficiency, Norwegian Computing Center,
can process calls to the timer module (symmetric multi- Forksningveien, 1B, Oslo 3, Pub. A9.
[19] A. A. Pritsker and P. J. Kiviat, Simulation with GASP-II. Englewood
processing), S. Glaser has pointed out that algorithms that Cliffs, NJ: Prentice-Hall, 1969.
tie up a common data structure for a large period of time [20] R. Souza et al., “GIGAswitch system: A high-performance packet-
will reduce efficiency. For instance, in Scheme 2, when switching platform,” Digital Tech. J., vol. 6, no. 1, Winter 1994.
[21] S. Szygenda, C. W. Hemming, and J. M. Hemphill, “Time flow
Processor A inserts a timer into the ordered list, other proces- mechanisms for use in digital logic simulations,” in Proc. 1971 Winter
sors cannot process timer module routines until Processor A Simulation Conf., New York.
[22] A. S. Tanenbaum, Computer Networks, 3rd ed. Upper Saddle River,
finishes and releases its semaphore. Schemes 5–7 seem suited NJ: Prentice-Hall, 1996.
for implementation in symmetric multiprocessors. However, [23] C. Thekkath, T. Nguyen, E. Moy, and E. Lazowska, “Implementing
given the recent research into techniques for maintaining network protocols at user level,” IEEE Trans. Networking, vol. 1, pp.
554–564, Oct. 1993.
consistency in multiprocessors, these differences may not be [24] E. Ulrich, “Time-sequenced logical simulation based on circuit delay
significant. and selective tracing of active network paths,” in 1965 ACM Nat. Conf.
834 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 5, NO. 6, DECEMBER 1997

[25] G. Varghese and A. Lauck, “Hashed and hierarchical timing wheels: Tony Lauck received the B.A. degree from Harvard University, Cambridge,
Data structures for the efficient implementation of a timer facility,” MA.
in Proc. 11th ACM Symp. Operating Syst. Principles, Nov. 1987, pp. He is currently an independent computer networking consultant. A former
171–180. Corporate Consulting Engineer at Digital Equipment Corporation, he was
[26] J. G. Vaucher and P. Duval, “A comparison of simulation event list responsible for Digital’s network architecture for 18 years.
algorithms,” Commun. ACM, vol. 18, 1975. Mr. Lauck was a member of the Internet Architecture Board (IAB) from
1990 to 1994.

George Varghese received the Ph.D. degree in

computer science from the Massachusetts Institute
of Technology (MIT), Cambridge, in 1992.
He worked from 1983 to 1993 at Digital design-
ing network protocols and doing systems research
as part of the DECNET architecture and advanced
development group. He has worked on design-
ing protocols and algorithms for DECNET and
GIGAswitch products. He is currently an Asso-
ciate Professor of Computer Science at Washington
University, St. Louis, MO, where he works on
distributed algorithms and efficient algorithms for network implementations.
Dr. Varghese has been awarded six patents with colleagues at DEC, with
six more patents being applied for. His Ph.D. dissertation on self-stablization
was jointly awarded the Sprowls Prize for best thesis in Computer Science
at MIT. He was among two computer scientists to receive the ONR Young
Investigator Award in 1996.