Messages in this thread | | | Date | Mon, 13 Apr 2009 12:57:50 -0700 (PDT) | From | Linus Torvalds <> | Subject | Re: [V4][PATCH 0/4]page fault retry with NOPAGE_RETRY |
| |
On Mon, 13 Apr 2009, Ying Han wrote: > > Benchmarks: > case 1. one application has a high count of threads each faulting in > different pages of a hugefile. Benchmark indicate that this double data > structure walking in case of major fault results in << 1% performance hit. > > case 2. add another thread in the above application which in a tight loop > of mmap()/munmap(). Here we measure loop count in the new thread while other > threads doing the same amount of work as case one. we got << 3% performance > hit on the Complete Time(benchmark value for case one) and 10% performance > improvement on the mmap()/munmap() counter. > > This patch helps a lot in cases we have writer which is waitting behind all > readers, so it could execute much faster.
Hmm. I normally think of "<<" as "much smaller than", but the way you use it makes me wonder. In particular, "<< 3%" sounds very odd. If it's much smaller than 3%, I'd have expected "<< 1%" again. So it probably isn't.
> benchmarks from Wufengguang: > Just tested the sparse-random-read-on-sparse-file case, and found the > performance impact to be 0.4% (8.706s vs 8.744s) in the worst case. > Kind of acceptable.
Well, have you tried the obvious optimization of _not_ doing the RETRY path when atomic_read(&mm->counter) == 1?
After all, if it's not a threaded app, and it doesn't have a possibility of concurrent mmap/fault, then why release the lock?
Linus
| |