188 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 7, NO. 2, APRIL 1999
The iSLIP Scheduling Algorithm
for Input-Queued Switches
Nick McKeown, Senior Member, IEEE
Abstract—An increasing number of high performance inter-
networking protocol routers, LAN and asynchronous transfer
mode (ATM) switches use a switched backplane based on a
crossbar switch. Most often, these systems use input queues
to hold packets waiting to traverse the switching fabric. It is
well known that if simple first in first out (FIFO) input queues
are used to hold packets then, even under benign conditions,
head-of-line (HOL) blocking limits the achievable bandwidth
to approximately 58.6% of the maximum. HOL blocking can
be overcome by the use of virtual output queueing, which is
described in this paper. A scheduling algorithm is used to con-
figure the crossbar switch, deciding the order in which packets
will be served. Recent results have shown that with a suitable
scheduling algorithm, 100% throughput can be achieved. In
this paper, we present a scheduling algorithm called
i
SLIP.
An iterative, round-robin algorithm,
i
SLIP can achieve 100%
throughput for uniform traffic, yet is simple to implement in
hardware. Iterative and noniterative versions of the algorithms
are presented, along with modified versions for prioritized traffic.
Simulation results are presented to indicate the performance
of
i
SLIP under benign and bursty traffic conditions. Prototype
and commercial implementations of
i
SLIP exist in systems with
aggregate bandwidths ranging from 50 to 500 Gb/s. When the
traffic is nonuniform,
i
SLIP quickly adapts to a fair scheduling
policy that is guaranteed never to starve an input queue. Finally,
we describe the implementation complexity of
i
SLIP. Based on
a two-dimensional (2-D) array of priority encoders, single-chip
schedulers have been built supporting up to 32 ports, and making
approximately 100 million scheduling decisions per second.
Index Terms— ATM switch, crossbar switch, input-queueing,
IP router, scheduling.
I. INTRODUCTION
I
N AN ATTEMPT to take advantage of the cell-switching
capacity of the asynchronous transfer mode (ATM), there
has recently been a merging of ATM switches and Inter-
net Protocol (IP) routers [29], [32]. This idea is already
being carried one step further, with cell switches forming
the core, or backplane, of high-performance IP routers [26],
[31], [6], [4]. Each of these high-speed switches and routers
is built around a crossbar switch that is configured using a
centralized scheduler, and each uses a fixed-size cell as a
transfer unit. Variable-length packets are segmented as they
arrive, transferred across the central switching fabric, and
then reassembled again into packets before they depart. A
crossbar switch is used because it is simple to implement and
Manuscript received November 19, 1996; revised February 9, 1998; ap-
proved by IEEE/ACM T
RANSACTIONS ON NETWORKING Editor H. J. Chao.
The author is with the Department of Electrical Engineering, Stanford
University, Stanford, CA 94305-9030 USA (e-mail: nickm@stanford.edu).
Publisher Item Identifier S 1063-6692(99)03593-1.
is nonblocking; it allows multiple cells to be transferred across
the fabric simultaneously, alleviating the congestion found on
a conventional shared backplane. In this paper, we describe an
algorithm that is designed to configure a crossbar switch using
a single-chip centralized scheduler. The algorithm presented
here attempts to achieve high throughput for best-effort unicast
traffic, and is designed to be simple to implement in hardware.
Our work was motivated by the design of two such systems:
the Cisco 12 000 GSR, a 50-Gb/s IP router, and the Tiny Tera:
a 0.5-Tb/s MPLS switch [7].
Before using a crossbar switch as a switching fabric, it is
important to consider some of the potential drawbacks; we
consider three here. First, the implementation complexity of an
-port crossbar switch increases with making crossbars
impractical for systems with a very large number of ports.
Fortunately, the majority of high-performance switches and
routers today have only a relatively small number of ports
(usually between 8 and 32). This is because the highest
performance devices are used at aggregation points where port
density is low.
1
Our work is, therefore, focussed on systems
with low port density. A second potential drawback of crossbar
switches is that they make it difficult to provide guaranteed
qualities of service. This is because cells arriving to the switch
must contend for access to the fabric with cells at both the
input and the output. The time at which they leave the input
queues and enter the crossbar switching fabric is dependent on
other traffic in the system, making it difficult to control when
a cell will depart. There are two common ways to mitigate
this problem. One is to schedule the transfer of cells from
inputs to outputs in a similar manner to that used in a time-
slot interchanger, providing peak bandwidth allocation for
reserved flows. This method has been implemented in at least
two commercial switches and routers.
2
The second approach
is to employ “speedup,” in which the core of the switch
runs faster than the connected lines. Simulation and analytical
results indicate that with a small speedup, a switch will deliver
cells quickly to their outgoing port, apparently independent of
contending traffic [27], [37]–[41]. While these techniques are
of growing importance, we restrict our focus in this paper to
the efficient and fast scheduling of best-effort traffic.
1
Some people believe that this situation will change in the future, and that
switches and routers with large aggregate bandwidths will support hundreds
or even thousands of ports. If these systems become real, then crossbar
switches—and the techniques that follow in this paper—may not be suitable.
However, the techniques described here will be suitable for a few years hence.
2
A peak-rate allocation method was supported by the DEC AN2
Gigaswitch/ATM [2] and the Cisco Systems LS2020 ATM Switch.
1063–6692/99$10.00 1999 IEEE
MCKEOWN: THE iSLIP SCHEDULING ALGORITHM FOR INPUT-QUEUED SWITCHES 189
Fig. 1. An input-queued switch with VOQ. Note that head of line blocking
is eliminated by using a separate queue for each output at each input.
A third potential drawback of crossbar switches is that they
(usually) employ input queues. When a cell arrives, it is placed
in an input queue where it waits its turn to be transferred across
the crossbar fabric. There is a popular perception that input-
queued switches suffer from inherently low performance due
to head-of-line (HOL) blocking. HOL blocking arises when
the input buffer is arranged as a single first in first out (FIFO)
queue: a cell destined to an output that is free may be held
up in line behind a cell that is waiting for an output that is
busy. Even with benign traffic, it is well known that HOL can
limit thoughput to just
[16]. Many techniques
have been suggested for reducing HOL blocking, for example
by considering the first
cells in the FIFO queue, where
[8], [13], [17]. Although these schemes can improve
throughput, they are sensitive to traffic arrival patterns and
may perform no better than regular FIFO queueing when the
traffic is bursty. But HOL blocking can be eliminated by using
a simple buffering strategy at each input port. Rather than
maintain a single FIFO queue for all cells, each input maintains
a separate queue for each output as shown in Fig. 1. This
scheme is called virtual output queueing (VOQ) and was first
introduced by Tamir et al. in [34]. HOL blocking is eliminated
because cells only queue behind cells that are destined to
the same output; no cell can be held up by a cell ahead of
it that is destined to a different output. When VOQ’s are
used, it has been shown possible to increase the throughput
of an input-queued switch from 58.6% to 100% for both
uniform and nonuniform traffic [25], [28]. Crossbar switches
that use VOQ’s have been employed in a number of studies
[1], [14], [19], [23], [34], research prototypes [26], [31], [33],
and commercial products [2], [6]. For the rest of this paper,
we will be considering crossbar switches that use VOQ’s.
When we use a crossbar switch, we require a scheduling
algorithm that configures the fabric during each cell time and
decides which inputs will be connected to which outputs; this
determines which of the
VOQ’s are served in each cell
time. At the beginning of each cell time, a scheduler examines
the contents of the
input queues and determines a conflict-
free match
between inputs and outputs. This is equivalent
to finding a bipartite matching on a graph with
vertices
[2], [25], [35]. For example, the algorithms described in [25]
and [28] that achieve 100% throughput, use maximum weight
bipartite matching algorithms [35], which have a running-time
complexity of
A. Maximum Size Matching
Most scheduling algorithms described previously are heuris-
tic algorithms that approximate a maximum size
3
matching
[1], [2], [5], [8], [18], [30], [36]. These algorithms attempt
to maximize the number of connections made in each cell
time, and hence, maximize the instantaneous allocation of
bandwidth. The maximum size matching for a bipartite graph
can be found by solving an equivalent network flow problem
[35]; we call the algorithm that does this maxsize. There exist
many maximum-size bipartite matching algorithms, and the
most efficient currently known converges in
time
[12].
4
The problem with this algorithm is that although it is
guaranteed to find a maximum match, for our application it
is too complex to implement in hardware and takes too long
to complete.
One question worth asking is “Does the maxsize algorithm
maximize the throughput of an input-queued switch?” The
answer is no; maxsize can cause some queues to be starved of
service indefinitely. Furthermore, when the traffic is nonuni-
form, maxsize cannot sustain very high throughput [25]. This is
because it does not consider the backlog of cells in the VOQ’s,
or the time that cells have been waiting in line to be served.
For practical high-performance systems, we desire algo-
rithms with the following properties.
• High Throughput: An algorithm that keeps the backlog
low in the VOQ’s; ideally, the algorithm will sustain an
offered load up to 100% on each input and output.
• Starvation Free: The algorithm should not allow a
nonempty VOQ to remain unserved indefinitely.
• Fast: To achieve the highest bandwidth switch, it is im-
portant that the scheduling algorithm does not become the
performance bottleneck; the algorithm should therefore
find a match as quickly as possible.
• Simple to Implement: If the algorithm is to be fast
in practice, it must be implemented in special-purpose
hardware, preferably within a single chip.
The
algorithm presented in this paper is designed
to meet these goals, and is currently implemented in a 16-
port commercial IP router with an aggregate bandwidth of 50
Gb/s [6], and a 32-port prototype switch with an aggregate
bandwidth of 0.5 Tb/s [26].
is based on the parallel
iterative matching algorithm (PIM) [2], and so to understand
its operation, we start by describing PIM. Then, in Section II,
we describe
and its performance. We then consider
some small modifications to
for various applications,
and finally consider its implementation complexity.
B. Parallel Iterative Matching
PIM was developed by DEC Systems Research Center for
the 16-port, 16 Gb/s AN2 switch [2].
5
Because it forms the
3
In some literature, the maximum size matching is called the maximum
cardinality matching or just the maximum bipartite matching.
4
This algorithm is equivalent to Dinic’s algorithm [9].
5
This switch was commercialized as the Gigaswitch/ATM.
190 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 7, NO. 2, APRIL 1999
(a)
(b) (c)
Fig. 2. An example of the three steps that make up one iteration of the PIM
scheduling algorithm [2]. In this example, the first iteration does not match
input 4 to output 4, even though it does not conflict with other connections.
This connection would be made in the second iteration. (a) Step 1: Request.
Each input makes a request to each output for which it has a cell. This is shown
here as a graph with all weights
w
ij
=1
:
(b) Step 2: Grant. Each output
selects an input uniformly among those that requested it. In this example,
inputs 1 and 3 both requested output 2. Output 2 chose to grant to input 3.
(c) Step 3: Accept. Each input selects an output uniformly among those that
granted to it. In this example, outputs 2 and 4 both granted to input 3. Input
3 chose to accept output 2.
basis of the algorithm described later, we will describe
the scheme in detail and consider some of its performance
characteristics.
PIM uses randomness to avoid starvation and reduce the
number of iterations needed to converge on a maximal-sized
match. A maximal-sized match (a type of on-line match) is
one that adds connections incrementally, without removing
connections made earlier in the matching process. In general, a
maximal match is smaller than a maximum-sized match, but is
much simpler to implement. PIM attempts to quickly converge
on a conflict-free maximal match in multiple iterations, where
each iteration consists of three steps. All inputs and outputs
are initially unmatched and only those inputs and outputs not
matched at the end of one iteration are eligible for matching in
the next. The three steps of each iteration operate in parallel on
each output and input and are shown in Fig. 2. The steps are:
Step 1: Request. Each unmatched input sends a request to
every output for which it has a queued cell.
Step 2: Grant. If an unmatched output receives any re-
quests, it grants to one by randomly selecting a request
uniformly over all requests.
Step 3: Accept. If an input receives a grant, it accepts one
by selecting an output randomly among those that granted to
this output.
By considering only unmatched inputs and outputs, each
iteration only considers connections not made by earlier iter-
ations.
Note that the independent output arbiters randomly select
a request among contending requests. This has three effects:
first, the authors in [2] show that each iteration will match
or eliminate, on average, at least
of the remaining pos-
sible connections, and thus, the algorithm will converge to a
maximal match, on average, in
iterations. Second,
it ensures that all requests will eventually be granted, ensuring
Fig. 3. Example of unfairness for PIM under heavy oversubscribed load with
more than one iterations. Because of the random and independent selection
by the arbiters, output 1 will grant to each input with probability 1/2, yet
input 1 will only accept output 1 a quarter of the time. This leads to different
rates at each output.
that no input queue is starved of service. Third, it means that
no memory or state is used to keep track of how recently a
connection was made in the past. At the beginning of each cell
time, the match begins over, independently of the matches that
were made in previous cell times. Not only does this simplify
our understanding of the algorithm, but it also makes analysis
of the performance straightforward; there is no time-varying
state to consider, except for the occupancy of the input queues.
Using randomness comes with its problems, however. First,
it is difficult and expensive to implement at high speed; each
arbiter must make a random selection among the members of
a time-varying set. Second, when the switch is oversubscribed,
PIM can lead to unfairness between connections. An extreme
example of unfairness for a 2
2 switch when the inputs are
oversubscribed is shown in Fig. 3. We will see examples later
for which PIM and some other algorithms are unfair when
no input or output is oversubscribed. Finally, PIM does not
perform well for a single iteration; it limits the throughput
to approximately 63%, only slightly higher than for a FIFO
switch. This is because the probability that an input will
remain ungranted is
hence as increases, the
throughput tends to
Although the algorithm
will often converge to a good match after several iterations,
the time to converge may affect the rate at which the switch
can operate. We would prefer an algorithm that performs well
with just a single iteration.
II. T
HE SLIP ALGORITHM WITH A SINGLE ITERATION
In this section we describe and evaluate the SLIP algorithm.
This section concentrates on the behavior of
SLIP with just
a single iteration per cell time. Later, we will consider
SLIP
with multiple iterations.
The
SLIP algorithm uses rotating priority (“round-robin”)
arbitration to schedule each active input and output in turn.
The main characteristic of
SLIP is its simplicity; it is readily
implemented in hardware and can operate at high speed. We
find that the performance of
SLIP for uniform traffic is
high; for uniform independent identically distributed (i.i.d.)
Bernoulli arrivals,
SLIP with a single iteration can achieve
100% throughput. This is the result of a phenomenon that we
encounter repeatedly; the arbiters in
SLIP have a tendency to
desynchronize with respect to one another.
A. Basic Round-Robin Matching Algorithm
SLIP is a variation of simple basic round-robin matching
algorithm (RRM). RRM is perhaps the simplest and most
MCKEOWN: THE iSLIP SCHEDULING ALGORITHM FOR INPUT-QUEUED SWITCHES 191
(a)
(b) (c)
Fig. 4. Example of the three steps of the RRM matching algorithm. (a) Step
1: Request. Each input makes a request to each output for which it has a cell.
Step 2: Grant. Each output selects the next requesting input at or after the
pointer in the round-robin schedule. Arbiters are shown here for outputs 2 and
4. Inputs 1 and 3 both requested output 2. Since
g
2
=1
;
output 2 grants to
input 1.
g
2
and
g
4
are updated to favor the input after the one that is granted.
(b) Step 3: Accept. Each input selects at most one output. The arbiter for input
1 is shown. Since
a
1
=1
;
input 1 accepts output 1.
a
1
is updated to point to
output 2. (c) When the arbitration is completed, a matching of size two has
been found. Note that this is less than the maximum sized matching of three.
obvious form of iterative round-robin scheduling algorithms,
comprising a 2-D array of round-robin arbiters; cells are
scheduled by round-robin arbiters at each output, and at each
input. As we shall see, RRM does not perform well, but it
helps us to understand how
SLIP performs, so we start here
with a description of RRM. RRM potentially overcomes two
problems in PIM: complexity and unfairness. Implemented as
priority encoders, the round-robin arbiters are much simpler
and can perform faster than random arbiters. The rotating
priority aids the algorithm in assigning bandwidth equally
and more fairly among requesting connections. The RRM
algorithm, like PIM, consists of three steps. As shown in
Fig. 4, for an
switch, each round-robin schedule
contains
ordered elements. The three steps of arbitration
are:
Step 1: Request. Each input sends a request to every output
for which it has a queued cell.
Step 2: Grant. If an output receives any requests, it chooses
the one that appears next in a fixed, roundrobin schedule
starting from the highest priority element. The output notifies
each input whether or not its request was granted. The pointer
to the highest priority element of the round-robin schedule is
incremented (modulo
to one location beyond the granted
input.
Step 3: Accept. If an input receives a grant, it accepts the
one that appears next in a fixed, round-robin schedule starting
from the highest priority element. The pointer
to the highest
priority element of the round-robin schedule is incremented
(modulo
to one location beyond the accepted output.
B. Performance of RRM for Bernoulli Arrivals
As an introduction to the performance of the RRM algo-
rithm, Fig. 5 shows the average delay as a function of offered
load for uniform independent and identically distributed (i.i.d.)
Fig. 5. Performance of RRM and
i
SLIP compared with PIM for i.i.d.
Bernoulli arrivals with destinations uniformly distributed over all outputs.
Results obtained using simulation for a 16
2
16 switch. The graph shows the
average delay per cell, measured in cell times, between arriving at the input
buffers and departing from the switch.
Fig. 6. 2
2
2 switch with RRM algorithm under heavy load. In the example
of Fig. 7, synchronization of output arbiters leads to a throughout of just 50%.
Bernoulli arrivals. For an offered load of just 63% RRM
becomes unstable.
6
The reason for the poor performance of RRM lies in the
rules for updating the pointers at the output arbiters. We
illustrate this with an example, shown in Fig. 6. Both inputs
1 and 2 are under heavy load and receive a new cell for
both outputs during every cell time. But because the output
schedulers move in lock-step, only one input is served during
each cell time. The sequence of requests, grants, and accepts
for four consecutive cell times are shown in Fig. 7. Note that
the grant pointers change in lock-step: in cell time 1, both
point to input 1, and during cell time 2, both point to input
2, etc. This synchronization phenomenon leads to a maximum
throughput of just 50% for this traffic pattern.
Synchronization of the grant pointers also limits perfor-
mance with random arrival patterns. Fig. 8 shows the number
of synchronized output arbiters as a function of offered load.
The graph plots the number of nonunique
’s, i.e., the number
of output arbiters that clash with another arbiter. Under low
6
The probability that an input will remain ungranted is
(
N
0
1
=N
)
N
;
hence as
N
increases, the throughput tends to
1
0
(1
=e
)
63%
:
192 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 7, NO. 2, APRIL 1999
Fig. 7. Illustration of low throughput for RRM caused by synchronization
of output arbiters. Note that pointers
[
g
i
]
stay synchronized, leading to a
maximum throughput of just 50%.
Fig. 8. Synchronization of output arbiters for RRM and
i
SLIP for i.i.d.
Bernoulli arrivals with destinations uniformly distributed over all outputs.
Results obtained using simulation for a 16
2
16 switch.
offered load, cells arriving for output will find in a
random position, equally likely to grant to any input. The
probability that
for all is
which for implies that the expected number of
arbiters with the same highest priority value is 9.9. This agrees
well with the simulation result for RRM in Fig. 8. As the
offered load increases, synchronized output arbiters tend to
move in lockstep and the degree of synchronization changes
only slightly.
III. T
HE SLIP ALGORITHM
The algorithm improves upon RRM by reducing the
synchronization of the output arbiters.
achieves this by
not moving the grant pointers unless the grant is accepted.
is identical to RRM except for a condition placed on
updating the grant pointers. The Grant step of RRM is changed
to:
Step 2: Grant. If an output receives any requests, it chooses
the one that appears next in a fixed round-robin schedule,
starting from the highest priority element. The output notifies
each input whether or not its request was granted. The pointer
to the highest priority element of the round-robin schedule
is incremented (modulo
to one location beyond the granted
input if, and only if, the grant is accepted in Step 3.
This small change to the algorithm leads to the following
properties of
with one iteration:
Property 1: Lowest priority is given to the most recently
made connection. This is because when the arbiters move their
pointers, the most recently granted (accepted) input (output)
becomes the lowest priority at that output (input). If input
successfully connects to output both and are updated
and the connection from input
to output becomes the lowest
priority connection in the next cell time.
Property 2: No connection is starved. This is because an
input will continue to request an output until it is successful.
The output will serve at most
other inputs first, waiting
at most
cell times to be accepted by each input. Therefore,
a requesting input is always served in less than
cell times.
Property 3: Under heavy load, all queues with a common
output have the same throughput. This is a consequence of
Property 2: the output pointer moves to each requesting input
in a fixed order, thus providing each with the same throughput.
But most importantly, this small change prevents the output
arbiters from moving in lock-step leading to a large improve-
ment in performance.
IV. S
IMULATED PERFORMANCE OF SLIP
A. With Benign Bernoulli Arrivals
Fig. 5 shows the performance improvement of
SLIP over
RRM. Under low load,
SLIP’s performance is almost identical
to RRM and FIFO; arriving cells usually find empty input
queues, and on average there are only a small number of inputs
requesting a given output. As the load increases, the number
of synchronized arbiters decreases (see Fig. 8), leading to a
large-sized match. In other words, as the load increases, we
can expect the pointers to move away from each, making it
more likely that a large match will be found quickly in the
next cell time. In fact, under uniform 100% offered load, the
SLIP arbiters adapt to a time-division multiplexing scheme,
providing a perfect match and 100% throughput. Fig. 9 is an
![Fig. 9. Illustration of 100% throughput foriSLIP caused by desynchronization of output arbiters. Note that pointers[gi] become desynchronized at the end of Cell 1 and stay desynchronized, leading to an alternating cycle of 2 cell times and a maximum throughput of 100%.](/figures/fig-9-illustration-of-100-throughput-forislip-caused-by-3uo2p4di.webp)
