From: Mark Fasheh Date: Sun, 5 Mar 2006 11:22:48 -0800 Subject: [Ocfs2-devel] Ocfs2 performance bugs of doom In-Reply-To: <20060303233617.51718c8e.akpm@osdl.org> References: <4408C2E8.4010600@google.com> <20060303233617.51718c8e.akpm@osdl.org> Message-ID: <20060305192247.GH31587@ca-server1.us.oracle.com> List-Id: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: Andrew Morton Cc: Daniel Phillips , linux-kernel@vger.kernel.org, ocfs2-devel@oss.oracle.com On Fri, Mar 03, 2006 at 11:36:17PM -0800, Andrew Morton wrote: > > - hash = full_name_hash(name, len); > > err, you might want to calculate that hash outside the spinlock. Hmm, good catch. > Maybe have a lock per bucket, too. That could be an avenue to explore. I guess we probably want to optimize the actual lookup and memory usage before looking at that. > A 1MB hashtable is verging on comical. How may data are there in total? Yes 1MB is much too big. We're looking right now at 3 lock resources per inode, so total number of elements depends upon your file system I suppose. Hopefully I'll have some time today to run some basic tests which may show us what size strikes a better balance of performance versus memory overhead. --Mark -- Mark Fasheh Senior Software Developer, Oracle mark.fasheh at oracle.com