From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from plane.gmane.org ([80.91.229.3]:53154 "EHLO plane.gmane.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753509Ab2JARAR (ORCPT ); Mon, 1 Oct 2012 13:00:17 -0400 Received: from list by plane.gmane.org with local (Exim 4.69) (envelope-from ) id 1TIjLk-0007XE-4U for linux-btrfs@vger.kernel.org; Mon, 01 Oct 2012 19:00:08 +0200 Received: from jfdmzpr03-ext.jf.intel.com ([134.134.139.72]) by main.gmane.org with esmtp (Gmexim 0.1 (Debian)) id 1AlnuQ-0007hv-00 for ; Mon, 01 Oct 2012 19:00:08 +0200 Received: from tim.c.chen by jfdmzpr03-ext.jf.intel.com with local (Gmexim 0.1 (Debian)) id 1AlnuQ-0007hv-00 for ; Mon, 01 Oct 2012 19:00:08 +0200 To: linux-btrfs@vger.kernel.org From: Tim Chen Subject: Re: [RFC][PATCH V2 0/4] Btrfs: introduce extent buffer cache to btrfs Date: Mon, 1 Oct 2012 16:53:19 +0000 (UTC) Message-ID: References: <5062E563.7090105@cn.fujitsu.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Sender: linux-btrfs-owner@vger.kernel.org List-ID: Miao Xie cn.fujitsu.com> writes: > > This patchset introduce extent buffer cache to btrfs. The basic idea > is to reduce the search time and the contentions of the extent buffer > lock by re-using the last search result. > > I ran stress.sh, xfstests and some other tools to test it, all of them > worked well. > > As a performance improvement patchset, of course we did performance test. > Because this patchset is to improve the b+ tree search, in other words, > it improves the performance of the metadata operations, we use file creation > test to measure it. So we ran 10 tasks, and all of them created 100000 files > in their own directories at the same time. As the result, we found this patchset > makes btrfs ~20% faster(98s -> 77s). > > we can pull this patchset from the URL > > git://github.com/miaoxie/linux-btrfs.git extent-buffer-cache > > Thanks > Miao > --- I tested the patchset with aim7's fileserver test with 16 processes on RAM emulated SCSI btrfs file system. I got about 18% speedup in throughput. The workload is a mix of file copy, read, write and file sync. The contention on btrfs_tree_lock operation is reduced with the patchset, by about 8.4% of cpu cycles (from 48.3% to 39.9%). Miao, I can send you the detailed profile if you are interested. Thanks. Tim Chen