Network measurement

Network collectors measure what a workload sends and receives. Application counters report payload; OS counters also capture packets and protocol overhead. Pick a collector that can attribute traffic to your workload before using a byte or packet limit.

Read the numbers

• Payload bytes — application data.

• Interface bytes — traffic at the interface boundary; includes protocol overhead and other processes.

• Packet — a unit of transmitted traffic; one message can become several packets.

• Drops — discarded traffic where observed.

• Latency — elapsed time for a declared operation.

• Throughput — completed operations or bytes per unit time.

High throughput does not imply low latency per request. Always read sent and received counters separately and check the scope before adding totals.

Attribution levels

• :network — counters the workload reports. CI-worthy for those counters.

• :network_interface — shared host interface. Not CI-worthy unless the machine is hermetic.

• :network_isolated — dedicated Linux network namespace. CI-worthy.

• measure_isolated_network_command — complete isolated process tree. CI-worthy, but covers the whole command lifecycle.

The default unprivileged namespace is loopback-only. It suits a service, database fixture, protocol implementation or client/server stack launched together. external_connectivity = true attaches an unprivileged slirp4netns TAP device for outbound IPv4 and DNS; the user-mode stack adds measurable overhead, which the capability manifest reports.

WSL2 process-tree capture

On Windows, PerfChecker can launch the same Linux namespace collector through WSL2. nftables counts packets; sysfs is a fallback.

using PerfChecker

isolation = NetworkIsolationSpec(provider = :wsl2_netns, distribution = "Ubuntu")
capabilities = network_isolation_capabilities(isolation; probe = true)
capabilities["supported"] || error(capabilities["reason"])

result = measure_isolated_network_command(
    ["julia", "--startup-file=no", "test/fixtures/network_loopback.jl"];
    spec = isolation, directory = pkgdir(PerfChecker), strict = true)

result.sample.packets_sent
result.sample.packets_received

CLI form:

julia --startup-file=no --project=. -e 'using PerfChecker; exit(perfchecker_main(ARGS))' -- network \
  --provider=linux_netns --output=perf/network.json -- \
  julia --startup-file=no perf/workload.jl

On Windows use --provider=wsl2_netns and optionally --distribution=Ubuntu. Add --external=true for outbound traffic (selects tap0 instead of lo). Output is omitted by default; opt in with --include-output=true. The result is perfchecker-isolated-network-result/1 with capture_layer = isolated_interface and attribution_scope = isolated_process_tree.

Interpretation

• TCP/IP framing, acknowledgements and retransmissions make wire bytes differ from payload bytes. Record both when it matters.

• On loopback, a transfer appears in both transmit and receive counters; summing them double-counts.

• A remote response time should include endpoint identity and experiment conditions; it is not an intrinsic package property.

• Windows host-interface collection is available through Get-NetAdapterStatistics. Native Windows process-tree attribution through ETW/WFP or Pktmon is not implemented; use the WSL provider.