EOS node performance bottleneck

General Discussion
21

Hi Cedric,

Really sorry for the nagging…

DId you have a chance to hear anything from your team about the scheduler? By the way, I did try out roundrobin, random etc for space.scheduler.type but there was no effect…

Do you think an upgrade from 5.3.10 to a newer version would be of any benefit?

22

Hello George,

We are still under discussion because it is not something easily fixable for now. A conclusion whether we can fix this or not has not been done yet.

Thanks for your patience,

Cedric

23

Thanks Cerdric.

@jleduc I remember that you suggested us to assign a seperate scheduling group to every EOS filesystem. For a single node that we were testing (with 8 FS), this means 8 scheduling groups. Does this mean that for 7 nodes (totalling 56 FS), we need to have 56 different scheduling groups?

24

Exactly: write scheduling in EOS is more or less round robin across scheduling groups (and also modulated by individual file system occupancy).

Therefore 1 scheduling group per fs for default space (archive) and same 1 scheduling group per fs for retrieve space gives you the best default spread upon write.

25

Many thanks for this Julien.