Linux standard OS tuning
pgbent’s Linux hosts re-apply a small set of OS settings after every boot via util/pgbent_tune, usually from util/newgis. None of these persist across reboot on their own (CPU governor and read_ahead_kb are runtime; NVMe coalescing is not saved to the drive). That is intentional: the same script is the checklist.
| Setting | What pgbent does | Why |
|---|---|---|
| CPU frequency governor | performance via cpupower | Avoid frequency scaling noise in timed runs |
| Block device read-ahead | blockdev --setra 4096 (→ read_ahead_kb=2048) | Buffered Postgres needs OS readahead; zero is a special case that starves miss paths |
| NVMe interrupt coalescing | nvme set-feature -f 8 -V 0 when nvme-cli is present | OEM defaults can hold every low-queue-depth read for 100 µs — see NVMe coalescing |
| TRIM | fstrim -av | Keep SSD free space honest between OSM cycles (can take a long time on multi-TB drives) |
Page-cache writeback (vm.dirty_*) is separate from this boot script — see Linux pdflush / writeback and the measured Linux dirty memory study.
Read-ahead
Classic Postgres lore ties readahead to sequential scans. That case was measured in Greg’s 9.0 PG book; public comparative work has been thin since. pgbent’s standard setra 4096 has been the lab default since its 2010 start.
Newer evidence on a scattered index→heap pattern (select-scatter-64, the Relations miniature):
- On court (NVIDIA DGX Spark),
read_ahead_kb=0roughly halved select-scatter-64 throughput vs 128/2048, both cold and hot, before and after the coalescing fix. 128 vs 2048 was flat. So “some readahead” matters even when the access pattern does not look like a seqscan; zero is the dangerous setting. - On twilight (x86), a public
read_ahead_kbladder (sets 134–139) shows the same shape: RA=0 soft (~¾ of 128+); 128 through 16384 flat. Contiguousselect-rangeon the same host (sets 140–145) is mostly flat across the ladder, with RA=0 only a mild dip at low concurrency.
OSM planet load (twilight, durable) closes the classic sequential / bulk-write story. Same face as the PG19 durable cell, but with read_ahead_kb=0 for the whole run (set 146) vs the lab default 2048 (set 121):
| cell | hours | OSM score | overall kNodes/s | index kNodes/s | rel/s |
|---|---|---|---|---|---|
| RA=2048 (121) | 5.43 | 1040 | 553 | 296 | 3849 |
| RA=0 (146) | 5.84 | 686 | 513 | 187 | 3440 |
Score drops by about a third. Most of the hit is index build (296→187); Relations only softens. On a fast NVMe x86 box, leaving RA at zero is not free for buffered Postgres loads.
Practical rule: do not leave Postgres volumes at array- or desktop-defaults of RA=0 because NTFS or interactive clients preferred that. Prefer on the order of pgbent’s 2048 KB unless you have measured otherwise. Isolation tables: ARM random read vs x86.
CPU governor
Timed benchmark and power comparisons assume the CPUs stay in performance. Mixing governors across a grid is a common false regression.
How this relates to Direct I/O / AIO
Oracle-oriented storage guides often assume Direct I/O, where the database owns prefetch and a dumb array is fine. Stock Postgres is still mostly on the buffered path: the OS readahead and completion behavior are part of the database’s miss latency. Until AIO and related work fully own that path on bare metal, these Linux knobs remain first-line checks — not trivia next to block alignment.