mirror of
https://github.com/zap-proto/bench.git
synced 2026-10-30 15:06:06 +00:00
zeekayandHanzo Dev
1b0d76a5de
"Repeat to the digit" overstated it at the large workload. Across the runs on file the allocations-per-request figure never varied at all, and the bytes-per-request figure spanned 131 264.8 to 131 268.9 at 64 KiB — four bytes in a hundred and thirty thousand, which is MemStats granularity. Wire sizes are byte-identical. The claim the harness rests on is that counts do not track load, and that is what the numbers support. Co-authored-by: Hanzo Dev <[email protected]>
64 lines
2.5 KiB
Bash
Executable File
64 lines
2.5 KiB
Bash
Executable File
#!/usr/bin/env bash
|
|
# Regenerate bench-results.txt. Everything the paper reports comes out of
|
|
# this one command, so a claim with no matching line here has no evidence.
|
|
#
|
|
# ./run.sh > bench-results.txt 2>&1
|
|
set -euo pipefail
|
|
cd "$(dirname "$0")"
|
|
|
|
# Allocation counts and wire sizes are counts, and they do not move with load:
|
|
# between load 99 and load 208 the allocations-per-request figure never varied
|
|
# at all and the bytes-per-request figure varied by four bytes in a hundred and
|
|
# thirty thousand. Throughput is a rate, and on a busy box it measures the
|
|
# scheduler. So the count sections run always and the rate sections run only on
|
|
# a quiet machine, and the file says which it contains. A number missing here
|
|
# was not measured; none is ever estimated.
|
|
cores=$(getconf _NPROCESSORS_ONLN 2>/dev/null || echo 1)
|
|
load=$(uptime | sed 's/.*averages*: *//' | tr -d ',' | awk '{print $1}')
|
|
# Quiet means most of the machine is idle: a one-minute load under two fifths
|
|
# of the core count. On the ten-core M1 Max this run was developed on that is
|
|
# a load of four.
|
|
limit=$(awk -v c="$cores" 'BEGIN{printf "%.1f", c*0.4}')
|
|
quiet=$(awk -v l="$load" -v m="$limit" 'BEGIN{print (l < m) ? "yes" : "no"}')
|
|
|
|
echo "=== environment ==="
|
|
go version
|
|
uname -sr
|
|
[ "$(uname -s)" = Darwin ] && sysctl -n machdep.cpu.brand_string hw.ncpu || true
|
|
uptime
|
|
echo "quiet enough for rates: $quiet (load $load, $cores cores, threshold $limit)"
|
|
go list -m all | grep -E 'zap-proto|fasthttp' || true
|
|
|
|
echo
|
|
echo "=== allocations, 3 repetitions ==="
|
|
go test -timeout=120m -run=TestMemoryPressure -v -count=3
|
|
|
|
echo
|
|
echo "=== wire bytes ==="
|
|
go test -timeout=120m -run=TestWireBytes -v -count=1
|
|
|
|
echo
|
|
if [ "$quiet" = yes ]; then
|
|
echo "=== concurrent throughput, 3 repetitions ==="
|
|
go test -timeout=120m -run=TestConcurrentThroughput -v -count=3
|
|
|
|
echo
|
|
echo "=== per-op benchmarks ==="
|
|
go test -timeout=120m -run=XXX -bench=. -benchmem -benchtime=2000x -count=3
|
|
else
|
|
echo "=== concurrent throughput: NOT MEASURED ==="
|
|
echo "Load average was $load on $cores cores when this run started. A rate"
|
|
echo "taken there is scheduler contention, not transport cost. Re-run on a"
|
|
echo "quiet machine to fill this section in."
|
|
echo
|
|
echo "=== per-op benchmarks: NOT MEASURED ==="
|
|
echo "Same reason. The B/op and allocs/op columns these would report are"
|
|
echo "already covered by the allocation section above, which is valid at"
|
|
echo "any load; only the ns/op column needs a quiet box, so the whole"
|
|
echo "section waits for one."
|
|
fi
|
|
|
|
echo
|
|
echo "=== environment after ==="
|
|
uptime
|