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
In 1.5.1 we introduced a feature that used the mmaphintMADV_WILLNEED when loading TSM files: #9380. The general idea was that because we need to scan TSM indexes on startup, using MADV_WILLNEED would let the kernel know it could read ahead pages, or at least prep them in the disk cache. Of course, we can't really control what the kernel does, just hint.
Due to #9534 (comment) we ended up having to remove the MADV_WILLNEED hint in 1.5.2. See #9614
Current behavior:
MADV_WILLNEED is not used as a hint to the kernel that we plan to scan the TSM index.
Desired behavior:
We have found that in some cases, users have moved from 1.5.1 (where we used MADV_WILLNEED) to >= 1.5.2 (where we don't), and have experienced significant IO Wait on >= 1.5.2. It's possible this is down to slow disks, and that in their case, on 1.5.1 their kernel is paging in TSM files in a way that is helping their IO performance.
We could consider making the MADV_WILLNEED hint a configuration option.
The text was updated successfully, but these errors were encountered:
Proposal:
In 1.5.1 we introduced a feature that used the
mmap
hintMADV_WILLNEED
when loading TSM files: #9380. The general idea was that because we need to scan TSM indexes on startup, usingMADV_WILLNEED
would let the kernel know it could read ahead pages, or at least prep them in the disk cache. Of course, we can't really control what the kernel does, just hint.Due to #9534 (comment) we ended up having to remove the
MADV_WILLNEED
hint in1.5.2
. See #9614Current behavior:
MADV_WILLNEED
is not used as a hint to the kernel that we plan to scan the TSM index.Desired behavior:
We have found that in some cases, users have moved from
1.5.1
(where we usedMADV_WILLNEED
) to>= 1.5.2
(where we don't), and have experienced significant IO Wait on>= 1.5.2
. It's possible this is down to slow disks, and that in their case, on1.5.1
their kernel is paging in TSM files in a way that is helping their IO performance.We could consider making the
MADV_WILLNEED
hint a configuration option.The text was updated successfully, but these errors were encountered: