ToplingDB dcompact bench results

← plain home

Updated (UTC): 2026-08-15 13:08:11+00:00

Latest dcompact run

run_id=31880805602 | result table | 2026-08-15 13:08:09+00:00

raw logs: ToplingDB zipkeyonly | ToplingDB zipkeyvalue | RocksDB v8.10 | RocksDB master (323d915d)

ToplingDB bench yaml: zipkeyonly db_bench-fillrandom.yaml | zipkeyonly db_bench-fillseq.yaml | zipkeyvalue db_bench-fillrandom.yaml | zipkeyvalue db_bench-fillseq.yaml

Bench configuration (dcompact)

Write-side db_bench CPU quota: 50% (from runner_env.txt cpu_quota_write). Topling compact: dcompact_worker — dcompact_worker runs outside the write cgroup and uses remaining CPU cores. RocksDB compact: compaction_service_spool — CompactionService spool + out-of-cgroup broker/worker on remaining cores.

Runner hardware/software

On-disk DB size (3.7GiB) ≤ block cache (7.8GiB) — cache can hold the entire dataset. So this benchmark mainly shows CPU and memory cost.

/dev/shm usage (disk space; after db_bench)

Allocated disk usage (IEC blocks). zipkeyonly does not compress values (speed-optimized). RocksDB uses per-level compression (L0 none, L1-L5 Snappy, L6 Zstd), corresponding to the ToplingDB zipkeyvalue variant's level_writers (lightweight upper levels, heavyweight L6). Space ratio = engine / v8.10. color sign: ToplingDB is Better/Worse.

Peak RSS (RAM; during db_bench)

RSS is Resident Set Size. RocksDB block cache = half physical memory (7.8GiB). Ratio = engine / v8.10. color sign: ToplingDB is Better/Worse.

RAM usage over time

Sampled once per second from /proc/statm plus open-file page cache. The bar below each plot shows stage name + duration (consecutive repeats merge, e.g. readseq×3); colored bands mark the same intervals. Stage names may overflow into a neighbor cell; a label under the bar is used only when names would overlap.

• rss: resident set size (pages currently in RAM for the process); rss = shared + anony.
• shared: shared resident pages; mostly readonly (cheap; OS prefers reclaiming these, no swap needed).
• anony: rss - shared; mostly readwrite anonymous pages (costly, needs swap).
• pagecache: kernel file page cache for regular files this process currently has open (cachestat(2) on /proc/pid/fd, deduped by inode; covers buffered read/write and mmap). Pages brought in only via buffered I/O are not charged to process RSS; mmap'd file pages may appear in both pagecache and RSS/shared, so the series can overlap. When reads hit this cache, they avoid disk I/O, and the DB benefits in its own way.
• anony+pc: anony+pagecache (sum of process anonymous RSS and open-file page cache; not a disjoint partition).

Comparison: db_bench fillrandom suite (perf)

Benchmarks: fillrandom, flush, compact, readseq×3, readrandom. RocksDB uses per-level compression (L0 none, L1-L5 Snappy, L6 Zstd), corresponding to the ToplingDB zipkeyvalue variant's level_writers (lightweight upper levels, heavyweight L6). compact row shows operations/time. color sign: ToplingDB is Better/Worse.

Comparison: db_bench fillseq suite (perf)

Same as fillrandom. RocksDB fillseq benefits from shortcuts: trivial_move on non-overlapping SSTs; refit level skips zstd on L6: faster, larger size. Seqno-zeroing compact still runs.

History

• 2026-08-15 13:08:09+00:00 — run_id=31880805602 — result table
• 2026-08-15 10:07:47+00:00 — run_id=31873910431 — result table
• 2026-08-14 15:45:19+00:00 — run_id=31800981467 — result table
• 2026-08-14 12:01:06+00:00 — run_id=31787128815 — result table
• 2026-08-09 17:28:01+00:00 — run_id=31324418704 — result table
• 2026-08-09 10:21:07+00:00 — run_id=31303589765 — result table
• 2026-08-09 09:29:35+00:00 — run_id=31298143542 — result table
• 2026-08-08 11:48:09+00:00 — run_id=31251963803 — result table