On 2017/12/15 3:21, Jaegeuk Kim wrote: > On 12/11, Chao Yu wrote: >> Hi Jaegeuk, >> >> On 2017/12/1 15:36, Jaegeuk Kim wrote: >>> Hi Chao, >>> >>> This is really hard to review and risky a lot to apply it shortly. Do we >>> have a >> >> I can understand your concern. >> >>> strong reason we have to do this? The original design goal was to minimize >>> allocation delay which is almost zero for now. Of course, I agreed that >>> there'd >> >> I agreed, For nid allocation performance there will be no room to improve >> since >> we have already used memory to exchange performance there. >> >> After applying this patch, I can encounter slight regression in scenario of >> continuous file creation. >> >> Anyway, goal of this patch is not aiming at performance, instead, it intends >> to >> reduce code complexity, and maybe, slightly improving performance of >> allocation >> in low-memory environment. So actually, this is a trade-off problem. >> >> If we decide to use old approach, I think what can be improved is that if we >> run >> out-of-memory, we can switch to nid bitmap searching in alloc_nid rather than >> preloading free nid entry in build_free_nid whose performance is memory >> sensitive. >> Thoughts? > > Hmm, in terms of performance, do we really suffer from memory pressure when > preloading free nids?
Sorry for late reply, I'm on trip now. I guess it will, let me emulate low memory environment, and do intensive allocation test, will report to you if there is any progress. Thanks,
