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
Using stat, a client can get a blocksize, which, according to man(2), is defined as
The st_blksize field gives the "preferred" blocksize for efficient file system I/O. (Writing to a file in smaller chunks may cause an inefficient read-modify-rewrite.)
Right now this field is independent of the actual blocksize (in terms of filestore) and equals 4KiB.
For example, this results in cat breaking one 1MiB request into smaller requests of size max(128KiB, blocksize) = 128KiB. This would be more efficient if we provided a client with blocksize of bs*64 (256KiB in case of the default bs).
The text was updated successfully, but these errors were encountered:
Successfully deployed PreferredBlockSizeMultiplier: 64. It had no effect on ior performance when using it without O_DIRECT. However, it indeed decreased number of read(2) syscalls while using cat on large files, as expected.
Also it had some benefitial impact on the dd/cp performance:
Using
stat
, a client can get a blocksize, which, according to man(2), is defined asRight now this field is independent of the actual blocksize (in terms of filestore) and equals 4KiB.
For example, this results in
cat
breaking one 1MiB request into smaller requests of sizemax(128KiB, blocksize) = 128KiB
. This would be more efficient if we provided a client with blocksize of bs*64 (256KiB in case of the default bs).The text was updated successfully, but these errors were encountered: