strawmANN benchmark report

A row flagged “… not server time” is one where the client saw far more than the server reported: below the rate at which a queue can form, what is left is the load generator's own scheduling.

The syscall rows count every descriptor, sockets included, so on a search row they measure the network rather than the disk.

measured via strawmANN: proc+cgroup / Qdrant: proc+cgroup. proc supplies syscall counts and block-layer bytes, cgroup supplies block-layer operations and bytes, so a row one interface does not carry reads n/a via …. unknown means it was not measured, and 0 means it was: an engine started with no --data-dir has no store, which is the row this section exists for. storage on disk is the level before the rows with a concurrent writer: during those an engine that rewrites segments is caught mid-rewrite, and the same row has read 3.47 and 10.11 GiB on two runs of one binary. peak RSS does include them, being a peak.

Per workload

A search row doing block-layer reads is an engine going to disk to answer a query.

Whether the engine was running while it ran. of wall is CPU over elapsed, waiting is runnable-but-not-scheduled, migrations checks the pinning claim, and switches gives voluntary over involuntary — the scheduler taking the core away against the engine choosing to sleep, which per unit of work is the cheapest signal of lock contention there is. A dash is not a zero: an index build's threads exit before the row does, and threads shows what fraction of the row the survivors account for.

waiting is runnable and not scheduled, blocked on disk is not runnable at all, and pressure is PSI for the engine's cgroup (some = at least one task stalled, full = every runnable task).

A dash under blocked on disk is not a zero: most hosts ship with kernel.task_delayacct=0, which reports the field as a permanent zero and would manufacture the strongest claim here — that the memory-resident engine never waits for a device — out of a sysctl. bench/setup.py apply turns it on.

What the core did per query, from perf stat attached to the engine for the same window as the /proc counters. §5's cost model is a set of claims about instructions per cycle, DRAM traffic and TLB reach; these are those quantities, measured on the row the headline quotes rather than on a microbenchmark.

Nothing here is scaled. Where perf had to multiplex the group, the values it prints are extrapolations from the fraction of the row each counter was on, and they are withheld rather than shown — the same rule the scheduler table applies to a thread that exited. An event this host does not implement is likewise blank, never zero.

The two instruments disagree on these rows. Qdrant: perf and /proc disagree on minor faults by up to 11% on W1, W2, W6-upload, W7-upload, W7-2bit-upload, W7-1p5bit-upload, W8-upload, W14-1bit-upload, W14-1p5bit-upload, W14-2bit-upload, W14-4bit-upload, W12-upload, W11-steady, W11; Qdrant: perf and /proc disagree on context switches by up to 23% on W8-upload, W14-1bit-upload, W14-1p5bit-upload, W14-2bit-upload, W14-4bit-upload. perf keeps an exiting thread's counts and the /proc sums do not, so a gap here is usually the same lost-thread effect the scheduler table reports as coverage.

Read back from the engine rather than taken from the request. --segments 1 sets Qdrant's default_segment_number, which its optimizer treats as a target, and the segment count is the largest confound in the ef comparison. strawmANN held 1 segment; Qdrant held 2 segments.

Where the engines disagree a cell reads strawmANN / Qdrant, and is marked.

After the mutating rows bench2 held strawmANN 1,250,000 (+250,000), 1,250,000 of them indexed; Qdrant 1,250,000 (+250,000), 1,246,100 of them indexed. The table above is the state the throughput, latency and recall rows searched; this is what W11 left, and is that row's subject rather than theirs.

bench1 was read back straight after its row and dropped, with no settle between, so an engine still building it reports a snapshot taken mid-ingest. Its cells are not marked as a disagreement.

Metric down, engine across, so a comparison is two adjacent cells. Metrics a row did not measure are dropped rather than shown empty.