vnlru: avoid cache line contention
ClosedPublic
Actions

Seems reasonable. Do you have any specific benchmark showing an actual improvement? Also, please add me as a reviewer for vnlru matters (or even anything related to the VFS).

olce accepted this revision.Tue, May 12, 9:33 AM

In D56951#1304662, @olce wrote:

Seems reasonable. Do you have any specific benchmark showing an actual improvement?

It depends on the cache line arrangement you would have gotten. With this and together with D56892 and D56949, we recovered a 5% performance regression on Spec SFS 2008, after unrelated changes caused object layout to change slightly. That was on platforms with dual Xeon 8260 (2 packages * 24 cores * 2 SMT) and other similar platforms.

But there may be zero improvement if the layout would not have been "bad".

Also, please add me as a reviewer for vnlru matters (or even anything related to the VFS).

Ack.

In D56951#1305103, @rlibby wrote:

It depends on the cache line arrangement you would have gotten. With this and together with D56892 and D56949, we recovered a 5% performance regression on Spec SFS 2008, after unrelated changes caused object layout to change slightly. That was on platforms with dual Xeon 8260 (2 packages * 24 cores * 2 SMT) and other similar platforms.

Ok. Wasn't aware of the existence of Spec SFS. Given the description of what the benchmarks are about (CIFS & NFS) and considering the work it takes to recycle a vnode with respect to the simple calculations you are optimizing here, I'm really surprised you can see as much as 5% difference.