[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Message-Id: <20240823173103.94978-1-jdamato@fastly.com>
Date: Fri, 23 Aug 2024 17:30:51 +0000
From: Joe Damato <jdamato@...tly.com>
To: netdev@...r.kernel.org
Cc: amritha.nambiar@...el.com,
sridhar.samudrala@...el.com,
sdf@...ichev.me,
peter@...eblog.net,
m2shafiei@...terloo.ca,
bjorn@...osinc.com,
hch@...radead.org,
willy@...radead.org,
willemdebruijn.kernel@...il.com,
skhawaja@...gle.com,
kuba@...nel.org,
Joe Damato <jdamato@...tly.com>,
Alexander Lobakin <aleksander.lobakin@...el.com>,
Alexander Viro <viro@...iv.linux.org.uk>,
bpf@...r.kernel.org (open list:BPF [MISC]:Keyword:(?:\b|_)bpf(?:\b|_)),
Breno Leitao <leitao@...ian.org>,
Christian Brauner <brauner@...nel.org>,
Daniel Borkmann <daniel@...earbox.net>,
"David S. Miller" <davem@...emloft.net>,
Eric Dumazet <edumazet@...gle.com>,
Jan Kara <jack@...e.cz>,
Jiri Pirko <jiri@...nulli.us>,
Johannes Berg <johannes.berg@...el.com>,
Jonathan Corbet <corbet@....net>,
linux-doc@...r.kernel.org (open list:DOCUMENTATION),
linux-fsdevel@...r.kernel.org (open list:FILESYSTEMS (VFS and infrastructure)),
linux-kernel@...r.kernel.org (open list),
Lorenzo Bianconi <lorenzo@...nel.org>,
Martin Karsten <mkarsten@...terloo.ca>,
Paolo Abeni <pabeni@...hat.com>,
Sebastian Andrzej Siewior <bigeasy@...utronix.de>
Subject: [PATCH net-next 0/6] Suspend IRQs during application busy periods
Greetings:
This series introduces a new mechanism, IRQ suspension, which allows
network applications using epoll to mask IRQs during periods of high
traffic while also reducing tail latency (compared to existing
mechanisms, see below) during periods of low traffic. In doing so, this
balances CPU consumption with network processing efficiency.
Martin Karsten (CC'd) and I have been collaborating on this series for
several months and have appreciated the feedback from the community on
our RFC [1]. We've updated the cover letter and kernel documentation in
an attempt to more clearly explain how this mechanism works, how
applications can use it, and how it compares to existing mechanisms in
the kernel. We've added an additional test case, 'fullbusy', achieved by
modifying libevent for comparison. See below for a detailed description,
link to the patch, and test results.
I briefly mentioned this idea at netdev conf 2024 (for those who were
there) and Martin described this idea in an earlier paper presented at
Sigmetrics 2024 [2].
~ The short explanation (TL;DR)
We propose adding a new sysfs parameter: irq_suspend_timeout to help
balance CPU usage and network processing efficiency when using IRQ
deferral and napi busy poll.
If this parameter is set to a non-zero value *and* a user application
has enabled preferred busy poll on a busy poll context (via the
EPIOCSPARAMS ioctl introduced in commit 18e2bf0edf4d ("eventpoll: Add
epoll ioctl for epoll_params")), then application calls to epoll_wait
for that context will cause device IRQs and softirq processing to be
suspended as long as epoll_wait successfully retrieves data from the
NAPI. Each time data is retrieved, the irq_suspend_timeout is deferred.
If/when network traffic subsides and epoll_wait returns no data, IRQ
suspension is immediately reverted back to the existing
napi_defer_hard_irqs and gro_flush_timeout mechanism which was
introduced in commit 6f8b12d661d0 ("net: napi: add hard irqs deferral
feature")).
The irq_suspend_timeout serves as a safety mechanism. If userland takes
a long time processing data, irq_suspend_timeout will fire and restart
normal NAPI processing.
For a more in depth explanation, please continue reading.
~ Comparison with existing mechanisms
Interrupt mitigation can be accomplished in napi software, by setting
napi_defer_hard_irqs and gro_flush_timeout, or via interrupt coalescing
in the NIC. This can be quite efficient, but in both cases, a fixed
timeout (or packet count) needs to be configured. However, a fixed
timeout cannot effectively support both low- and high-load situations:
At low load, an application typically processes a few requests and then
waits to receive more input data. In this scenario, a large timeout will
cause unnecessary latency.
At high load, an application typically processes many requests before
being ready to receive more input data. In this case, a small timeout
will likely fire prematurely and trigger irq/softirq processing, which
interferes with the application's execution. This causes overhead, most
likely due to cache contention.
While NICs attempt to provide adaptive interrupt coalescing schemes,
these cannot properly take into account application-level processing.
An alternative packet delivery mechanism is busy-polling, which results
in perfect alignment of application processing and network polling. It
delivers optimal performance (throughput and latency), but results in
100% cpu utilization and is thus inefficient for below-capacity
workloads.
We propose to add a new packet delivery mode that properly alternates
between busy polling and interrupt-based delivery depending on busy and
idle periods of the application. During a busy period, the system
operates in busy-polling mode, which avoids interference. During an idle
period, the system falls back to interrupt deferral, but with a small
timeout to avoid excessive latencies. This delivery mode can also be
viewed as an extension of basic interrupt deferral, but alternating
between a small and a very large timeout.
This delivery mode is efficient, because it avoids softirq execution
interfering with application processing during busy periods. It can be
used with blocking epoll_wait to conserve cpu cycles during idle
periods. The effect of alternating between busy and idle periods is that
performance (throughput and latency) is very close to full busy polling,
while cpu utilization is lower and very close to interrupt mitigation.
~ Usage details
IRQ suspension is introduced via a sysfs parameter that controls the
maximum time that IRQs can be suspended.
Here's how it is intended to work:
- An administrator sets the existing sysfs parameters for
napi_defer_hard_irqs and gro_flush_timeout to enable IRQ deferral.
- An administrator sets the new sysfs parameter irq_suspend_timeout
to a larger value than gro_flush_timeout to enable IRQ suspension.
- The user application issues the existing epoll ioctl to set the
prefer_busy_poll flag on the epoll context.
- The user application then calls epoll_wait to busy poll for network
events, as it normally would.
- If epoll_wait returns events to userland, IRQs are suspended for the
duration of irq_suspend_timeout.
- If epoll_wait finds no events and the thread is about to go to
sleep, IRQ handling using napi_defer_hard_irqs and gro_flush_timeout
is resumed.
As long as epoll_wait is retrieving events, IRQs (and softirq
processing) for the NAPI being polled remain disabled. When network
traffic reduces, eventually a busy poll loop in the kernel will retrieve
no data. When this occurs, regular IRQ deferral using gro_flush_timeout
for the polled NAPI is re-enabled.
Unless IRQ suspension is continued by subsequent calls to epoll_wait, it
automatically times out after the irq_suspend_timeout timer expires.
Regular deferral is also immediately re-enabled when the epoll context
is destroyed.
~ Usage scenario
The target scenario for IRQ suspension as packet delivery mode is a
system that runs a dominant application with substantial network I/O.
The target application can be configured to receive input data up to a
certain batch size (via epoll_wait maxevents parameter) and this batch
size determines the worst-case latency that application requests might
experience. Because packet delivery is suspended during the target
application's processing, the batch size also determines the worst-case
latency of concurrent applications.
gro_flush_timeout should be set as small as possible, but large enough to
make sure that a single request is likely not being interfered with.
irq_suspend_timeout is largely a safety mechanism against misbehaving
applications. It should be set large enough to cover the processing of an
entire application batch, i.e., the factor between gro_flush_timeout and
irq_suspend_timeout should roughly correspond to the maximum batch size
that the target application would process in one go.
~ Design rationale
The implementation of the IRQ suspension mechanism very nicely dovetails
with the existing mechanism for IRQ deferral when preferred busy poll is
enabled (introduced in commit 7fd3253a7de6 ("net: Introduce preferred
busy-polling"), see that commit message for more details).
The existing IRQ deferral mechanism works together with our proposal,
and as such, it seems natural to put irq_suspend_time at the same level
of napi_defer_hard_irqs and gro_flush_timeout as a per-device sysfs
parameter with the hope that these parameters will eventually be
migrated to per-napi settings.
While it would be possible to inject the suspend timeout via
the existing epoll ioctl, it is more natural to avoid this path for two
reasons:
1. Using a sysfs parameter ensures admin oversight for using the
mechanism and its configuration based on overall system objectives
2. An epoll context is linked to NAPI IDs as file descriptors are added;
this means any epoll context might suddenly be associated with a
different net_device if the application were to replace all existing
fds with fds from a different device. In this case, the scope of the
suspend timeout becomes unclear and many edge cases for both the user
application and the kernel are introduced
Only a single iteration through napi busy polling is needed for this
mechanism to work effectively. Since an important objective for this
mechanism is preserving cpu cycles, exactly one iteration of the napi
busy loop is invoked when busy_poll_usecs is set to 0.
~ Important call outs in the implementation
- Enabling per epoll-context preferred busy poll will now effectively
lead to a nonblocking iteration through napi_busy_loop, even when
busy_poll_usecs is 0. See patch 4.
- Patches apply cleanly on commit d785ed945de6 ("net: wwan: t7xx: PCIe
reset rescan"). When commit b4988e3bd1f0 ("eventpoll: Annotate
data-race of busy_poll_usecs") is picked up from the vfs folks, there
will be a very minor merge conflict with patch 4. patch 4 already
includes the fix picked up vfs, but modifies the same line by adding an
extra conditional.
- In the future, time permitting, I hope to enable support for
napi_defer_hard_irqs, gro_flush_timeout (introduced in commit
6f8b12d661d0 ("net: napi: add hard irqs deferral feature")), and
irq_suspend_timeout (introduced in this series) on a per-NAPI basis
(presumably via netdev-genl).
~ Benchmark configs & descriptions
The changes were benchmarked with memcached [3] using the benchmarking
tool mutilate [4].
To facilitate benchmarking, a small patch [5] was applied to memcached
1.6.29 (the latest memcached release as of this submission) to allow
setting per-epoll context preferred busy poll and other settings via
environment variables. Another small patch [6] was applied to libevent
to enable full busy-polling.
Multiple scenarios were benchmarked as described below and the scripts
used for producing these results can be found on github [7] (note: all
scenarios use NAPI-based traffic splitting via SO_INCOMING_ID by passing
-N to memcached):
- base:
- no other options enabled
- deferX:
- set defer_hard_irqs to 100
- set gro_flush_timeout to X,000
- napibusy:
- set defer_hard_irqs to 100
- set gro_flush_timeout to 200,000
- enable busy poll via the existing ioctl (busy_poll_usecs = 64,
busy_poll_budget = 64, prefer_busy_poll = true)
- fullbusy:
- set defer_hard_irqs to 100
- set gro_flush_timeout to 5,000,000
- enable busy poll via the existing ioctl (busy_poll_usecs = 1000,
busy_poll_budget = 64, prefer_busy_poll = true)
- change memcached's nonblocking epoll_wait invocation (via
libevent) to using a 1 ms timeout
- suspendX:
- set defer_hard_irqs to 100
- set gro_flush_timeout to X,000
- set irq_suspend_timeout to 20,000,000
- enable busy poll via the existing ioctl (busy_poll_usecs = 0,
busy_poll_budget = 64, prefer_busy_poll = true)
~ Benchmark results
Tested on:
Single socket AMD EPYC 7662 64-Core Processor
Hyperthreading disabled
4 NUMA Zones (NPS=4)
16 CPUs per NUMA zone (64 cores total)
2 x Dual port 100gbps Mellanox Technologies ConnectX-5 Ex EN NIC
The test machine is configured such that a single interface has 8 RX
queues. The queues' IRQs and memcached are pinned to CPUs that are
NUMA-local to the interface which is under test. The NIC's interrupt
coalescing configuration is left at boot-time defaults.
Results:
Results are shown below. The mechanism added by this series is
represented by the 'suspend' cases. Data presented shows a summary over
10 runs of each test case [8] using the scripts on github [7]. For,
latency the median of the 10 runs is shown. For throughput and CPU
utilization, the average is shown.
These results were captured using the scripts on github [7] to
illustrate how this approach compares with other pre-existing
mechanisms. This data is not to be interpreted as scientific data
captured in a fully isolated lab setting, but instead as best effort,
illustrative information comparing and contrasting tradeoffs.
Compare:
- Throughput (MAX) and latencies of base vs suspend.
- CPU usage of napibusy and fullbusy during lower load (200K, 400K for
example) vs suspend.
- Latency of the defer variants vs suspend as timeout and load
increases.
The overall takeaway is that the suspend variants provide a superior
combination of high throughput, low latency, and low cpu utilization
compared to all other variants. Each of the suspend variants works very
well, but some fine-tuning between latency and cpu utilization is still
possible by tuning the small timeout (gro_flush_timeout).
base
load qps avglat 95%lat 99%lat cpu
200K 200042 113 234 416 28
400K 399983 141 271 730 42
600K 599951 162 386 712 64
800K 800057 350 1151 2278 84
1000K 962651 4168 5980 7611 99
MAX 982241 4426 5809 7317 99
defer10
load qps avglat 95%lat 99%lat cpu
200K 200025 53 119 140 31
400K 400064 59 131 156 53
600K 599930 72 146 204 76
800K 800002 656 3236 5455 97
1000K 887890 4806 6288 8568 99
MAX 914970 4863 6008 6361 100
defer20
load qps avglat 95%lat 99%lat cpu
200K 199971 59 122 149 26
400K 400029 67 139 167 46
600K 599985 78 152 195 68
800K 799931 250 1009 2481 89
1000K 896278 4662 5876 6422 99
MAX 966120 4549 5652 6076 100
defer50
load qps avglat 95%lat 99%lat cpu
200K 200005 78 133 175 22
400K 400000 88 159 192 39
600K 600067 97 173 214 58
800K 799933 177 519 1302 81
1000K 948784 4378 6078 8941 98
MAX 1015637 4358 5354 6365 100
defer200
load qps avglat 95%lat 99%lat cpu
200K 200025 164 255 304 18
400K 399964 182 276 331 34
600K 599924 204 303 361 48
800K 800013 253 401 727 73
1000K 979727 3380 5690 7648 98
MAX 1022960 4322 5598 6652 100
fullbusy
load qps avglat 95%lat 99%lat cpu
200K 200000 47 113 129 100
400K 400071 51 123 148 100
600K 599942 58 128 176 100
800K 799966 65 138 191 100
1000K 1000099 83 167 233 100
MAX 1159578 3797 4281 4371 100
napibusy
load qps avglat 95%lat 99%lat cpu
200K 200022 100 238 272 56
400K 399897 77 223 273 83
600K 600077 65 158 255 96
800K 800007 74 147 237 99
1000K 999992 88 173 242 100
MAX 1067116 4060 5903 10231 99
suspend10
load qps avglat 95%lat 99%lat cpu
200K 199984 51 119 139 32
400K 400077 56 127 153 51
600K 599944 64 137 190 69
800K 800012 71 145 200 84
1000K 1000031 95 183 300 94
MAX 1146106 3852 4313 4462 100
suspend20
load qps avglat 95%lat 99%lat cpu
200K 199966 57 120 146 29
400K 400016 61 133 158 47
600K 599997 68 141 187 65
800K 800057 77 153 210 81
1000K 1000059 107 191 397 92
MAX 1150835 3813 4284 4373 100
suspend50
load qps avglat 95%lat 99%lat cpu
200K 199940 72 125 170 25
400K 399981 76 145 184 42
600K 600054 82 157 201 58
800K 799940 93 175 352 75
1000K 1000000 108 200 269 89
MAX 1139291 3834 4330 4383 100
suspend200
load qps avglat 95%lat 99%lat cpu
200K 199983 149 250 298 19
400K 399968 155 270 326 35
600K 599982 159 285 348 51
800K 800079 162 297 361 67
1000K 999944 175 311 392 84
MAX 1128863 3891 4334 4383 100
~ FAQ
- Can the new timeout value be threaded through the new epoll ioctl ?
Only with difficulty. The epoll ioctl sets options on an epoll
context and the NAPI ID associated with an epoll context can change
based on what file descriptors a user app adds to the epoll context.
This would introduce complexity in the API from the user perspective
and also complexity in the kernel.
This new sysfs parameter, which is similar to and used in
combination with napi_defer_hard_irqs and gro_flush_timeout should
be exposed to users in the same way: via sysfs.
At some point in the future, per-NAPI support for
napi_defer_hard_irqs, gro_flush_timeout, and irq_suspend_timeout
(presumably via netdev-genl) can be added.
- Can irq suspend be built by combining NIC coalescing and
gro_flush_timeout ?
No. The problem is that the long timeout must engage if and only if
prefer-busy is active.
When using NIC coalescing for the short timeout (without
napi_defer_hard_irqs/gro_flush_timeout), an interrupt after an idle
period will trigger softirq, which will run napi polling. At this
point, prefer-busy is not active, so NIC interrupts would be
re-enabled. Then it is not possible for the longer timeout to
interject to switch control back to polling. In other words, only by
using the software timer for the short timeout, it is possible to
extend the timeout without having to reprogram the NIC timer or
reach down directly and disable interrupts.
Using gro_flush_timeout for the long timeout also has problems, for
the same underlying reason. In the current napi implementation,
gro_flush_timeout is not tied to prefer-busy. We'd either have to
change that and in the process modify the existing deferral
mechanism, or introduce a state variable to determine whether
gro_flush_timeout is used as long timeout for irq suspend or whether
it is used for its default purpose. In an earlier version, we did
try something similar to the latter and made it work, but it ends up
being a lot more convoluted than our current proposal.
- Isn't it already possible to combine busy looping with irq deferral?
Yes, in fact enabling irq deferral via napi_defer_hard_irqs and
gro_flush_timeout is a precondition for prefer_busy_poll to have an
effect. If the application also uses a tight busy loop with
essentially nonblocking epoll_wait (accomplished with a very short
timeout parameter), this is the fullbusy case shown in the results.
An application using blocking epoll_wait is shown as the napibusy
case in the result. It's a hybrid approach that provides limited
latency benefits compared to the base case and plain irq deferral,
but not as good as fullbusy or suspend.
~ Special thanks
Several people were involved in earlier stages of the development of this
mechanism whom we'd like to thank:
- Peter Cai (CC'd), for the initial kernel patch and his contributions
to the paper.
- Mohammadamin Shafie (CC'd), for testing various versions of the kernel
patch and providing helpful feedback.
Thanks,
Martin and Joe
[1]: https://lore.kernel.org/netdev/20240812125717.413108-1-jdamato@fastly.com/
[2]: https://doi.org/10.1145/3626780
[3]: https://github.com/memcached/memcached/blob/master/doc/napi_ids.txt
[4]: https://github.com/leverich/mutilate
[5]: https://raw.githubusercontent.com/martinkarsten/irqsuspend/main/patches/memcached.patch
[6]: https://raw.githubusercontent.com/martinkarsten/irqsuspend/main/patches/libevent.patch
[7]: https://github.com/martinkarsten/irqsuspend
[8]: https://github.com/martinkarsten/irqsuspend/tree/main/results
rfc -> v1:
- Cover letter updated to include more details.
- Patch 1 updated to remove the documentation added. This was moved to
patch 6 with the rest of the docs (see below).
- Patch 5 updated to fix an error uncovered by the kernel build robot.
See patch 5's changelog for more details.
- Patch 6 added which updates kernel documentation.
Joe Damato (1):
docs: networking: Describe irq suspension
Martin Karsten (5):
net: Add sysfs parameter irq_suspend_timeout
net: Suspend softirq when prefer_busy_poll is set
net: Add control functions for irq suspension
eventpoll: Trigger napi_busy_loop, if prefer_busy_poll is set
eventpoll: Control irq suspension for prefer_busy_poll
Documentation/networking/napi.rst | 112 +++++++++++++++++++++++++++++-
fs/eventpoll.c | 34 ++++++++-
include/linux/netdevice.h | 2 +
include/net/busy_poll.h | 3 +
net/core/dev.c | 53 ++++++++++++--
net/core/net-sysfs.c | 18 +++++
6 files changed, 213 insertions(+), 9 deletions(-)
--
2.25.1
Powered by blists - more mailing lists