From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from eggs.gnu.org ([2001:4830:134:3::10]:60362) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1cuBis-0000Rg-TJ for qemu-devel@nongnu.org; Sat, 01 Apr 2017 01:37:15 -0400 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1cuBip-0007XV-Q3 for qemu-devel@nongnu.org; Sat, 01 Apr 2017 01:37:14 -0400 Received: from mx1.redhat.com ([209.132.183.28]:33896) by eggs.gnu.org with esmtps (TLS1.0:DHE_RSA_AES_256_CBC_SHA1:32) (Exim 4.71) (envelope-from ) id 1cuBip-0007V8-Hf for qemu-devel@nongnu.org; Sat, 01 Apr 2017 01:37:11 -0400 Date: Sat, 1 Apr 2017 13:37:07 +0800 From: Fam Zheng Message-ID: <20170401053707.GJ11195@lemon.lan> References: MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: Subject: Re: [Qemu-devel] Performance problem and improvement about block drive on NFS shares with libnfs List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: Jaden Liang Cc: qemu-devel@nongnu.org On Sat, 04/01 13:23, Jaden Liang wrote: > Hello, > > I ran qemu with drive file via libnfs recently, and found some performance > problem and an improvement idea. > > I started qemu with 6 drives parameter like nfs://127.0.0.1/dir/vm-disk-x.qcow2 > which linked to a local NFS server, then used iometer in guest machine to test > the 4K random read or random write IO performance. I found that while the IO > depth go up, the IOPS hit a bottleneck. I looked into the causes, found that the > main thread of qemu used 100% CPU. From the perf data, it show the CPU heats are > send / recv calls in libnfs. By reading the source code of libnfs and qemu block > drive of nfs.c, libnfs only support single work thread, and the network events > of nfs interface in qemu are all registered in the epoll of main thread. That is > the cause why main thread uses 100% CPU. > > After the analysis above, there is an improvement idea comes up. I start a > thread for every drive while libnfs open drive file, then create an epoll in > every drive thread to handle all of the network events. I have finished an demo > modification in block/nfs.c, then rerun iometer in the guest machine, the > performance increased a lot. Random read IOPS increases almost 100%, random > write IOPS increases about 68%. > > Test model details > VM configure: 6 vdisks in 1 VM > Test tool and parameter: iometer with 4K random read and randwrite > Backend physical drive: 2 SSDs, 6 vdisks are seperated in 2 SSDs > > Before modified: > IO Depth 1 2 4 8 16 32 > 4K randread 16659 28387 42932 46868 52108 55760 > 4K randwrite 12212 19456 30447 30574 35788 39015 > > After modified: > IO Depth 1 2 4 8 16 32 > 4K randread 17661 33115 57138 82016 99369 109410 > 4K randwrite 12669 21492 36017 51532 61475 65577 > > I could put a up to coding standard patch later. Now I want to get some advise > about this modification. Is this a reasonable solution to improve performance in > NFS shares? Or there is another better way? > > Any suggestions would be great! Also please feel free to ask question. Just one comment: in block/file-posix.c (aio=threads), there is a thread pool that does something similar, using the code util/thread-pool.c. Maybe it's usable for your block/nfs.c change too. Also a question: have you considered modifying libnfs to create more worker threads? That way all applications using libnfs can benefit. Fam