Messages in this thread | | | Subject | Re: [MM] Make mm counters per cpu instead of atomic | From | "Zhang, Yanmin" <> | Date | Tue, 24 Nov 2009 16:02:33 +0800 |
| |
On Mon, 2009-11-23 at 08:31 -0600, Christoph Lameter wrote: > On Mon, 23 Nov 2009, Zhang, Yanmin wrote: > > > Another theoretic issue is below scenario: > > Process A get the read lock on cpu 0 and is scheduled to cpu 2 to unlock. Then > > it's scheduled back to cpu 0 to repeat the step. eventually, the reader counter > > will overflow. Considering multiple thread cases, it might be faster to > > overflow than what we imagine. When it overflows, processes will hang there. > > True.... We need to find some alternative to per cpu data to scale mmap > sem then. I ran lots of benchmarks such like specjbb2005/hackbench/tbench/dbench/iozone /sysbench_oltp(mysql)/aim7 against percpu tree(based on 2.6.32-rc7) on a 4*8*2 logical cpu machine, and didn't find big result difference between with your patch and without your patch.
| |