From mboxrd@z Thu Jan 1 00:00:00 1970 From: Hubert Bahr Subject: extremely long delays are they possible? Date: Sat, 16 Oct 2010 01:02:12 -0500 Message-ID: <4CB93FE4.6030005@hbahr.org> Reply-To: hab@hbahr.org Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit To: linux-ext4@vger.kernel.org Return-path: Received: from hrndva-omtalb.mail.rr.com ([71.74.56.125]:61760 "EHLO hrndva-omtalb.mail.rr.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752026Ab0JPGCN (ORCPT ); Sat, 16 Oct 2010 02:02:13 -0400 Sender: linux-ext4-owner@vger.kernel.org List-ID: In reality I would prefer writes only occur to an ssd if the FS write cache is overflowing. The application is package rebuilding. The package is brought in from nfs expanded built (compiled etc) then the results written back to an nfs mount and the log written to Cluster Head then everything erased. However, if there are build problems the expanded file system and results are left intact. My Memory Size is large enough (16GB) to buffer the whole process except when a build problem occur's. Since this can be hours into a build I would prefer to capture. This is part of a cluster batch process that is only examined at the expected end of the total batch process. Capturing the problems precludes using only a ramfs, but I would like to avoid needless writes to the ssd. I am afraid that embedded fsyncs may blow this anyway. Maybe I could use some on error mv but I prefer the dynamic memory sharing rather than a fixed ramfs. The longest single package build I have observed is 11.5 hours and it also took the most space. would fstab options like ext4 commit=50000,noatime,delalloc,data=writeback give me the desired action? Thanks Hubert