You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
PXB-3862 : Handle OS-cached pages in the storage probe
On a buffered server the probe's sample regions may already sit in the
OS page cache (warmed by the server or a previous scan), and a cached
read times RAM, not the device: on a fully cached tablespace the round
trip measured 1us and the read request cost collapsed to the 64KiB
floor, understating how much merging the storage deserves.
The probe already dropped its samples with posix_fadvise(DONTNEED), but
newer kernels cache a sequentially read file in blocks of up to 2MB
(XFS since kernel 5.17, ext4 since 6.13) and ignore the advice unless
it covers a whole block, so the page-sized evictions did nothing there.
Evict a whole 2MB-aligned window around every sample instead, right
before reading it. The probe gives up ~40MB of cache on a multi-GB
file; the copy itself still reads a warm file from RAM. Verified
against a fully cached 21GB tablespace: calibration now reports device
numbers (178us round trip) while the backup reads only 79MB from the
device.
0 commit comments