|
Add a benchmarks to demonstrate the performance cliff for local_storage
get as the number of local_storage maps increases beyond current
local_storage implementation's cache size.
"sequential get" and "interleaved get" benchmarks are added, both of
which do many bpf_task_storage_get calls on sets of task local_storage
maps of various counts, while considering a single specific map to be
'important' and counting task_storage_gets to the important map
separately in addition to normal 'hits' count of all gets. Goal here is
to mimic scenario where a particular program using one map - the
important one - is running on a system where many other local_storage
maps exist and are accessed often.
While "sequential get" benchmark does bpf_task_storage_get for map 0, 1,
..., {9, 99, 999} in order, "interleaved" benchmark interleaves 4
bpf_task_storage_gets for the important map for every 10 map gets. This
is meant to highlight performance differences when important map is
accessed far more frequently than non-important maps.
A "hashmap control" benchmark is also included for easy comparison of
standard bpf hashmap lookup vs local_storage get. The benchmark is
similar to "sequential get", but creates and uses BPF_MAP_TYPE_HASH
instead of local storage. Only one inner map is created - a hashmap
meant to hold tid -> data mapping for all tasks. Size of the hashmap is
hardcoded to my system's PID_MAX_LIMIT (4,194,304). The number of these
keys which are actually fetched as part of the benchmark is
configurable.
Addition of this benchmark is inspired by conversation with Alexei in a
previous patchset's thread [0], which highlighted the need for such a
benchmark to motivate and validate improvements to local_storage
implementation. My approach in that series focused on improving
performance for explicitly-marked 'important' maps and was rejected
with feedback to make more generally-applicable improvements while
avoiding explicitly marking maps as important. Thus the benchmark
reports both general and important-map-focused metrics, so effect of
future work on both is clear.
Regarding the benchmark results. On a powerful system (Skylake, 20
cores, 256gb ram):
Hashmap Control
===============
num keys: 10
hashmap (control) sequential get: hits throughput: 20.900 ± 0.334 M ops/s, hits latency: 47.847 ns/op, important_hits throughput: 20.900 ± 0.334 M ops/s
num keys: 1000
hashmap (control) sequential get: hits throughput: 13.758 ± 0.219 M ops/s, hits latency: 72.683 ns/op, important_hits throughput: 13.758 ± 0.219 M ops/s
num keys: 10000
hashmap (control) sequential get: hits throughput: 6.995 ± 0.034 M ops/s, hits latency: 142.959 ns/op, important_hits throughput: 6.995 ± 0.034 M ops/s
num keys: 100000
hashmap (control) sequential get: hits throughput: 4.452 ± 0.371 M ops/s, hits latency: 224.635 ns/op, important_hits throughput: 4.452 ± 0.371 M ops/s
num keys: 4194304
hashmap (control) sequential get: hits throughput: 3.043 ± 0.033 M ops/s, hits latency: 328.587 ns/op, important_hits throughput: 3.043 ± 0.033 M ops/s
Local Storage
=============
num_maps: 1
local_storage cache sequential get: hits throughput: 47.298 ± 0.180 M ops/s, hits latency: 21.142 ns/op, important_hits throughput: 47.298 ± 0.180 M ops/s
local_storage cache interleaved get: hits throughput: 55.277 ± 0.888 M ops/s, hits latency: 18.091 ns/op, important_hits throughput: 55.277 ± 0.888 M ops/s
num_maps: 10
local_storage cache sequential get: hits throughput: 40.240 ± 0.802 M ops/s, hits latency: 24.851 ns/op, important_hits throughput: 4.024 ± 0.080 M ops/s
local_storage cache interleaved get: hits throughput: 48.701 ± 0.722 M ops/s, hits latency: 20.533 ns/op, important_hits throughput: 17.393 ± 0.258 M ops/s
num_maps: 16
local_storage cache sequential get: hits throughput: 44.515 ± 0.708 M ops/s, hits latency: 22.464 ns/op, important_hits throughput: 2.782 ± 0.044 M ops/s
local_storage cache interleaved get: hits throughput: 49.553 ± 2.260 M ops/s, hits latency: 20.181 ns/op, important_hits throughput: 15.767 ± 0.719 M ops/s
num_maps: 17
local_storage cache sequential get: hits throughput: 38.778 ± 0.302 M ops/s, hits latency: 25.788 ns/op, important_hits throughput: 2.284 ± 0.018 M ops/s
local_storage cache interleaved get: hits throughput: 43.848 ± 1.023 M ops/s, hits latency: 22.806 ns/op, important_hits throughput: 13.349 ± 0.311 M ops/s
num_maps: 24
local_storage cache sequential get: hits throughput: 19.317 ± 0.568 M ops/s, hits latency: 51.769 ns/op, important_hits throughput: 0.806 ± 0.024 M ops/s
local_storage cache interleaved get: hits throughput: 24.397 ± 0.272 M ops/s, hits latency: 40.989 ns/op, important_hits throughput: 6.863 ± 0.077 M ops/s
num_maps: 32
local_storage cache sequential get: hits throughput: 13.333 ± 0.135 M ops/s, hits latency: 75.000 ns/op, important_hits throughput: 0.417 ± 0.004 M ops/s
local_storage cache interleaved get: hits throughput: 16.898 ± 0.383 M ops/s, hits latency: 59.178 ns/op, important_hits throughput: 4.717 ± 0.107 M ops/s
num_maps: 100
local_storage cache sequential get: hits throughput: 6.360 ± 0.107 M ops/s, hits latency: 157.233 ns/op, important_hits throughput: 0.064 ± 0.001 M ops/s
local_storage cache interleaved get: hits throughput: 7.303 ± 0.362 M ops/s, hits latency: 136.930 ns/op, important_hits throughput: 1.907 ± 0.094 M ops/s
num_maps: 1000
local_storage cache sequential get: hits throughput: 0.452 ± 0.010 M ops/s, hits latency: 2214.022 ns/op, important_hits throughput: 0.000 ± 0.000 M ops/s
local_storage cache interleaved get: hits throughput: 0.542 ± 0.007 M ops/s, hits latency: 1843.341 ns/op, important_hits throughput: 0.136 ± 0.002 M ops/s
Looking at the "sequential get" results, it's clear that as the
number of task local_storage maps grows beyond the current cache size
(16), there's a significant reduction in hits throughput. Note that
current local_storage implementation assigns a cache_idx to maps as they
are created. Since "sequential get" is creating maps 0..n in order and
then doing bpf_task_storage_get calls in the same order, the benchmark
is effectively ensuring that a map will not be in cache when the program
tries to access it.
For "interleaved get" results, important-map hits throughput is greatly
increased as the important map is more likely to be in cache by virtue
of being accessed far more frequently. Throughput still reduces as #
maps increases, though.
To get a sense of the overhead of the benchmark program, I
commented out bpf_task_storage_get/bpf_map_lookup_elem in
local_storage_bench.c and ran the benchmark on the same host as the
'real' run. Results:
Hashmap Control
===============
num keys: 10
hashmap (control) sequential get: hits throughput: 54.288 ± 0.655 M ops/s, hits latency: 18.420 ns/op, important_hits throughput: 54.288 ± 0.655 M ops/s
num keys: 1000
hashmap (control) sequential get: hits throughput: 52.913 ± 0.519 M ops/s, hits latency: 18.899 ns/op, important_hits throughput: 52.913 ± 0.519 M ops/s
num keys: 10000
hashmap (control) sequential get: hits throughput: 53.480 ± 1.235 M ops/s, hits latency: 18.699 ns/op, important_hits throughput: 53.480 ± 1.235 M ops/s
num keys: 100000
hashmap (control) sequential get: hits throughput: 54.982 ± 1.902 M ops/s, hits latency: 18.188 ns/op, important_hits throughput: 54.982 ± 1.902 M ops/s
num keys: 4194304
hashmap (control) sequential get: hits throughput: 50.858 ± 0.707 M ops/s, hits latency: 19.662 ns/op, important_hits throughput: 50.858 ± 0.707 M ops/s
Local Storage
=============
num_maps: 1
local_storage cache sequential get: hits throughput: 110.990 ± 4.828 M ops/s, hits latency: 9.010 ns/op, important_hits throughput: 110.990 ± 4.828 M ops/s
local_storage cache interleaved get: hits throughput: 161.057 ± 4.090 M ops/s, hits latency: 6.209 ns/op, important_hits throughput: 161.057 ± 4.090 M ops/s
num_maps: 10
local_storage cache sequential get: hits throughput: 112.930 ± 1.079 M ops/s, hits latency: 8.855 ns/op, important_hits throughput: 11.293 ± 0.108 M ops/s
local_storage cache interleaved get: hits throughput: 115.841 ± 2.088 M ops/s, hits latency: 8.633 ns/op, important_hits throughput: 41.372 ± 0.746 M ops/s
num_maps: 16
local_storage cache sequential get: hits throughput: 115.653 ± 0.416 M ops/s, hits latency: 8.647 ns/op, important_hits throughput: 7.228 ± 0.026 M ops/s
local_storage cache interleaved get: hits throughput: 138.717 ± 1.649 M ops/s, hits latency: 7.209 ns/op, important_hits throughput: 44.137 ± 0.525 M ops/s
num_maps: 17
local_storage cache sequential get: hits throughput: 112.020 ± 1.649 M ops/s, hits latency: 8.927 ns/op, important_hits throughput: 6.598 ± 0.097 M ops/s
local_storage cache interleaved get: hits throughput: 128.089 ± 1.960 M ops/s, hits latency: 7.807 ns/op, important_hits throughput: 38.995 ± 0.597 M ops/s
num_maps: 24
local_storage cache sequential get: hits throughput: 92.447 ± 5.170 M ops/s, hits latency: 10.817 ns/op, important_hits throughput: 3.855 ± 0.216 M ops/s
local_storage cache interleaved get: hits throughput: 128.844 ± 2.808 M ops/s, hits latency: 7.761 ns/op, important_hits throughput: 36.245 ± 0.790 M ops/s
num_maps: 32
local_storage cache sequential get: hits throughput: 102.042 ± 1.462 M ops/s, hits latency: 9.800 ns/op, important_hits throughput: 3.194 ± 0.046 M ops/s
local_storage cache interleaved get: hits throughput: 126.577 ± 1.818 M ops/s, hits latency: 7.900 ns/op, important_hits throughput: 35.332 ± 0.507 M ops/s
num_maps: 100
local_storage cache sequential get: hits throughput: 111.327 ± 1.401 M ops/s, hits latency: 8.983 ns/op, important_hits throughput: 1.113 ± 0.014 M ops/s
local_storage cache interleaved get: hits throughput: 131.327 ± 1.339 M ops/s, hits latency: 7.615 ns/op, important_hits throughput: 34.302 ± 0.350 M ops/s
num_maps: 1000
local_storage cache sequential get: hits throughput: 101.978 ± 0.563 M ops/s, hits latency: 9.806 ns/op, important_hits throughput: 0.102 ± 0.001 M ops/s
local_storage cache interleaved get: hits throughput: 141.084 ± 1.098 M ops/s, hits latency: 7.088 ns/op, important_hits throughput: 35.430 ± 0.276 M ops/s
Adjusting for overhead, latency numbers for "hashmap control" and
"sequential get" are:
hashmap_control_1k: ~53.8ns
hashmap_control_10k: ~124.2ns
hashmap_control_100k: ~206.5ns
sequential_get_1: ~12.1ns
sequential_get_10: ~16.0ns
sequential_get_16: ~13.8ns
sequential_get_17: ~16.8ns
sequential_get_24: ~40.9ns
sequential_get_32: ~65.2ns
sequential_get_100: ~148.2ns
sequential_get_1000: ~2204ns
Clearly demonstrating a cliff.
In the discussion for v1 of this patch, Alexei noted that local_storage
was 2.5x faster than a large hashmap when initially implemented [1]. The
benchmark results show that local_storage is 5-10x faster: a
long-running BPF application putting some pid-specific info into a
hashmap for each pid it sees will probably see on the order of 10-100k
pids. Bench numbers for hashmaps of this size are ~10x slower than
sequential_get_16, but as the number of local_storage maps grows far
past local_storage cache size the performance advantage shrinks and
eventually reverses.
When running the benchmarks it may be necessary to bump 'open files'
ulimit for a successful run.
[0]: https://lore.kernel.org/all/20220420002143.1096548-1-davemarchevsky@fb.com
[1]: https://lore.kernel.org/bpf/20220511173305.ftldpn23m4ski3d3@MBP-98dd607d3435.dhcp.thefacebook.com/
Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Link: https://lore.kernel.org/r/20220620222554.270578-1-davemarchevsky@fb.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
|
|
While working on BPF ringbuf implementation, testing, and benchmarking, I've
developed a pretty generic and modular benchmark runner, which seems to be
generically useful, as I've already used it for one more purpose (testing
fastest way to trigger BPF program, to minimize overhead of in-kernel code).
This patch adds generic part of benchmark runner and sets up Makefile for
extending it with more sets of benchmarks.
Benchmarker itself operates by spinning up specified number of producer and
consumer threads, setting up interval timer sending SIGALARM signal to
application once a second. Every second, current snapshot with hits/drops
counters are collected and stored in an array. Drops are useful for
producer/consumer benchmarks in which producer might overwhelm consumers.
Once test finishes after given amount of warm-up and testing seconds, mean and
stddev are calculated (ignoring warm-up results) and is printed out to stdout.
This setup seems to give consistent and accurate results.
To validate behavior, I added two atomic counting tests: global and local.
For global one, all the producer threads are atomically incrementing same
counter as fast as possible. This, of course, leads to huge drop of
performance once there is more than one producer thread due to CPUs fighting
for the same memory location.
Local counting, on the other hand, maintains one counter per each producer
thread, incremented independently. Once per second, all counters are read and
added together to form final "counting throughput" measurement. As expected,
such setup demonstrates linear scalability with number of producers (as long
as there are enough physical CPU cores, of course). See example output below.
Also, this setup can nicely demonstrate disastrous effects of false sharing,
if care is not taken to take those per-producer counters apart into
independent cache lines.
Demo output shows global counter first with 1 producer, then with 4. Both
total and per-producer performance significantly drop. The last run is local
counter with 4 producers, demonstrating near-perfect scalability.
$ ./bench -a -w1 -d2 -p1 count-global
Setting up benchmark 'count-global'...
Benchmark 'count-global' started.
Iter 0 ( 24.822us): hits 148.179M/s (148.179M/prod), drops 0.000M/s
Iter 1 ( 37.939us): hits 149.308M/s (149.308M/prod), drops 0.000M/s
Iter 2 (-10.774us): hits 150.717M/s (150.717M/prod), drops 0.000M/s
Iter 3 ( 3.807us): hits 151.435M/s (151.435M/prod), drops 0.000M/s
Summary: hits 150.488 ± 1.079M/s (150.488M/prod), drops 0.000 ± 0.000M/s
$ ./bench -a -w1 -d2 -p4 count-global
Setting up benchmark 'count-global'...
Benchmark 'count-global' started.
Iter 0 ( 60.659us): hits 53.910M/s ( 13.477M/prod), drops 0.000M/s
Iter 1 (-17.658us): hits 53.722M/s ( 13.431M/prod), drops 0.000M/s
Iter 2 ( 5.865us): hits 53.495M/s ( 13.374M/prod), drops 0.000M/s
Iter 3 ( 0.104us): hits 53.606M/s ( 13.402M/prod), drops 0.000M/s
Summary: hits 53.608 ± 0.113M/s ( 13.402M/prod), drops 0.000 ± 0.000M/s
$ ./bench -a -w1 -d2 -p4 count-local
Setting up benchmark 'count-local'...
Benchmark 'count-local' started.
Iter 0 ( 23.388us): hits 640.450M/s (160.113M/prod), drops 0.000M/s
Iter 1 ( 2.291us): hits 605.661M/s (151.415M/prod), drops 0.000M/s
Iter 2 ( -6.415us): hits 607.092M/s (151.773M/prod), drops 0.000M/s
Iter 3 ( -1.361us): hits 601.796M/s (150.449M/prod), drops 0.000M/s
Summary: hits 604.849 ± 2.739M/s (151.212M/prod), drops 0.000 ± 0.000M/s
Benchmark runner supports setting thread affinity for producer and consumer
threads. You can use -a flag for default CPU selection scheme, where first
consumer gets CPU #0, next one gets CPU #1, and so on. Then producer threads
pick up next CPU and increment one-by-one as well. But user can also specify
a set of CPUs independently for producers and consumers with --prod-affinity
1,2-10,15 and --cons-affinity <set-of-cpus>. The latter allows to force
producers and consumers to share same set of CPUs, if necessary.
Signed-off-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20200512192445.2351848-3-andriin@fb.com
|
Dave Marchevsky
Hou Tao
Joanne Koong
Andrii Nakryiko