I/O Performance Monitoring
 

Hello all,

I'm researching and I'm looking for understanding of what is happening in terms of writing data to a filesystem mounted as NFS.

Here is what I see.

• 4 separate threads start writing a 1GB file each to the /nfstest filesystem.
• /proc/meminfo (Dirty) begins to increase
• At the same time I see /proc/slabinfo (nfs_page) start to increase (both num_objs and active_objs)
• /proc/meminfo (Dirty) reaches ~2GB and then /proc/meminfo (Writeback) begins to increase (elt ~7 seconds)

It seems at this point slabinfo/nfs_page is writing to disk and, at times, will accept more pages from meminfo/Writeback when it can (if I'm saying this wrong please correct me).

What I notice at the end is Dirty and Writeback are done (only about 70kB) while slabinfo/nfs_page continues writing. When slabinfo/nfs_page (both num_objs and active_objs) finally finishes writing the 4 threads complete.

I'm just beginning to understand this so forgive my newbiness.

What I'd like to understand is how meminfo/Dirty|Writeback work with slabinfo/nfs_page. If there are ways for me to know how much slab can accept prior to starting to write and if it can continue to accept new objs while writing the objs currently in slab (if I'm even saying it right).

Can anyone shed any light on this?

I'm more than happy to read about all of this, but I have had trouble finding info on slab and how it works.

Thanks,
Michael Cunningham

Slabs are just a means of optimizing frequent memory allocation requests. A good overview exists here. It's a little old now, but the concepts are valid.
Nowadays, the SLUB allocator is usually used by current kernels to manage slabs - basically better, faster.

I saw something a while back about why NFS slabs - can't find it at the moment.

Thanks a lot syg00. I'll look at the article and search for nfs slabs as well.

Michael Cunningham