bench: 90th percentile latency, per-connection link emulation

wrk's and mix's 90th percentile is recorded (p90_ms) and shown in the
report.py --latency tables when present.

--flow-rate limits each connection between the client and the proxy
(fq maxrate), under --netem when both are given, now with a large netem
queue. Both directions are shaped in the client's namespace, the
incoming one through an ifb, so the emulation runs on the client's CPUs;
segmentation offload and GRO are off on that link so that netem drops
single packets.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Vladimir DubrovinandClaude Opus 5.5 committed 2026-10-14 11:02:30 +03:00
1 parent b1a551c5b1
commit 7140ffb43d
4 files changed
+65 -20

No files matched your search

+11 -8
View File
@@ -27,9 +27,10 @@ of the group whose range (the average of its slowest runs to that of its
fastest) reaches the best one's; * marks the best of the row, both groups
together, in the same way. --split puts the splice group in a table of its own
(* still compares across both), --prefix connect- takes only the scenarios so
named. --latency makes one table of the runs' 50th and 99th percentile
latency for a single CSV instead, ranked by the 99th. --overall marks the best of the row only, in bold and with *: for
tables where one group has a single proxy, its group's best says nothing.
named. --latency makes one table of the runs' 50th, 90th and 99th
percentile latency for a single CSV instead, ranked by the 99th. --overall
marks the best of the row only, in bold and with *: for tables where one
group has a single proxy, its group's best says nothing.
"""
import argparse
@@ -237,8 +238,8 @@ def combined(machines, labels, splice, prefix=None, split=False, overall_only=Fa
def latency(path, labels, splice, prefix=None):
"""Latency of one machine's runs: each cell the medians of the runs' 50th
and 99th percentiles, in milliseconds. Bold: the lowest 99th percentile
"""Latency of one machine's runs: each cell the medians of the runs' 50th,
90th (when recorded) and 99th percentiles, in milliseconds. Bold: the lowest 99th percentile
of its group and any other whose runs overlap it; *: the same over the
row. Scenarios without a latency (iperf3 streams) are left out."""
by, proxies, scenarios = {}, [], []
@@ -269,7 +270,9 @@ def latency(path, labels, splice, prefix=None):
rule.append(":-:")
head += [labels.get(p, p) for p in g]
rule += ["---:"] * len(g)
out = ["| scenario, p50 / p99 ms | " + " | ".join(head) + " |",
pcts = ["p50_ms", "p90_ms", "p99_ms"] if any(values(rs, "p90_ms") for rs in by.values()) \
else ["p50_ms", "p99_ms"]
out = [f"| scenario, {' / '.join(c[:3] for c in pcts)} ms | " + " | ".join(head) + " |",
"|---|" + "|".join(rule) + "|"]
for s in scenarios:
overall = winners(by, s, everyone, "p99_ms", False)
@@ -283,7 +286,7 @@ def latency(path, labels, splice, prefix=None):
if not sel:
cells.append("–")
continue
text = f"{ms(median(sel, 'p50_ms'))} / {ms(median(sel, 'p99_ms'))}"
text = " / ".join(ms(median(sel, c)) for c in pcts)
if p in best:
text = f"**{text}**"
if p in overall:
@@ -312,7 +315,7 @@ def main():
ap.add_argument("--splice", action="append", default=[],
help="a proxy (as named in the CSV) using splice(); repeat")
ap.add_argument("--latency", action="store_true",
help="one table of latency, 50th / 99th percentile, for the "
help="one table of latency, 50th / 90th / 99th percentile, for the "
"one CSV given")
args = ap.parse_args()
labels = dict(l.split("=", 1) for l in args.label)