Tune a buffer pool
A database keeps recently used disk pages in a buffer pool in memory, so a page read twice is read from disk once. When the pool is full, a replacement policy decides which page to evict. How the pool performs depends far more on the workload than on the policy's sophistication. What should this pool do?
You will end with a replacement policy, the pool's usable size under load, and the settings that matter more than either.
Before you choose
Use declared access patterns if your engine has them
If operators can tell the pool "this is a one-time scan": use that. It beats every adaptive policy, because the operator knows what the policy can only guess. Check whether your engine exposes it before tuning any policy.
Decide
What dominates the workload?
Why
A scan touches every page once, so no replacement policy can keep useful pages while it runs: the scan evicts them. Point lookups on skewed keys reuse a small hot set, so even a small pool captures most of the benefit.
Where you can land
Isolate scans in a ring buffer
Use it when scans dominate.
Give each scan a small private ring of pages, so it cannot evict the shared hot set.
Clock
Use it when many cores share the pool, or the workload is point lookups.
Approximates least-recently-used with one reference bit per page and almost no work on a hit.
Before you size it: the hit-rate curve flattens quickly under skew, so compute the hit rate the next gigabyte buys before buying it.
2Q or ARC
Use it when scans and point lookups mix, and core count is modest.
Separates pages seen once from pages seen repeatedly, so a scan cannot flush the frequently used ones.
What it costs: extra metadata and more work per hit.
Whichever you land on
Find the usable size under load
Usable size = configured size − pinned pages − dirty pages not yet written. That, not the configured number, is what memory rationing must be given.
Check the cleaner keeps up with writes
If dirty pages accumulate faster than the background writer flushes them: the pool fills with pages it cannot evict, and reads stall. The fix is upstream of the replacement policy: write throughput, flush settings.
Back the pool with huge pages, and turn transparent huge pages off
Explicit huge pages cut address-translation misses for a large pool; the kernel's automatic version causes stalls. It is a yes-or-no setting with a large effect, and the one most often left at its default.