On 06/29/2011 08:27 PM, Avi Kivity wrote: > On 06/29/2011 03:28 PM, Xiao Guangrong wrote: >> On 06/29/2011 08:18 PM, Avi Kivity wrote: >> > On 06/29/2011 02:50 PM, Xiao Guangrong wrote: >> >> >> > >> >> >> > I think we should do this unconditionally. The cost of ping-ponging the shared cache line containing reader_counter will increase with large smp counts. On the other hand, zap_page is very rare, so it can be a little slower. Also, less code paths = easier to understand. >> >> >> > >> >> >> >> >> >> On soft mmu, zap_page is very frequently, it can cause performance regression in my test. >> >> > >> >> > Any idea what the cause of the regression is? It seems to me that simply deferring freeing shouldn't have a large impact. >> >> > >> >> >> >> I guess it is because the page is freed too frequently, i have done the test, it shows >> >> about 3219 pages is freed per second >> >> >> >> Kernbench performance comparing: >> >> >> >> the origin way: 3m27.723 >> >> free all shadow page in rcu context: 3m30.519 >> > >> > I don't recall seeing such a high free rate. Who is doing all this zapping? >> > >> > You may be able to find out with the function tracer + call graph. >> > >> >> I looked into it before, it is caused by "write flood" detected, i also noticed >> some pages are zapped and allocation again and again, maybe we need to improve >> the algorithm of detecting "write flood". > > Ok. Let's drop the two paths, and put this improvement on the TODO instead. > Avi, i am sorry, i do not understand it clearly, it means keep the patch as the original way and do the improvement after it merged? -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
