From: "Martin K. Petersen" Subject: Re: Tuning for Compact Flash Date: Mon, 29 Sep 2008 11:17:49 -0400 Message-ID: References: <48DE2D20.1090703@csr.com> <48DFA291.2030206@hp.com> <48E0C8DA.6090502@weinigel.se> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: jim owens , linux-fsdevel@vger.kernel.org To: Christer Weinigel Return-path: Received: from rgminet01.oracle.com ([148.87.113.118]:19741 "EHLO rgminet01.oracle.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751729AbYI2PSM (ORCPT ); Mon, 29 Sep 2008 11:18:12 -0400 In-Reply-To: <48E0C8DA.6090502@weinigel.se> (Christer Weinigel's message of "Mon\, 29 Sep 2008 14\:23\:54 +0200") Sender: linux-fsdevel-owner@vger.kernel.org List-ID: >>>>> "Christer" == Christer Weinigel writes: Christer> Of course it won't behave like a hard drive, but it would be Christer> nice if there was a file system with performance that didn't Christer> suck as badly on a compact flash. Well, you are limited by the capabilities of the FTL. And as far as CF goes the FTL is usually pretty braindead. Making sure things are properly aligned really helps on "real" SSDs as well as 4KB hw sector drives. I'm working on some code that exposes these parameters so we can perform that alignment when creating filesystems. But as far as CF is concerned I have tested many and not found a single one that is suitable for a regular (random) filesystem write workload. To the extent that for embedded devices I ended up putting a root filesystem tarball on the CF and unpacking it into ramfs during boot. And just to give you an idea: I'm getting two orders of magnitude less random write ops on a (high-performance) CF than on a flash SSD from the same vendor. In both cases with I/O aligned to the flash page boundary. -- Martin K. Petersen Oracle Linux Engineering