From mboxrd@z Thu Jan 1 00:00:00 1970 From: Tejun Heo Subject: Re: Performance issue with SATA port multiplier Date: Thu, 24 Feb 2011 16:21:52 +0100 Message-ID: <20110224152152.GR7840@htj.dyndns.org> References: Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Return-path: Received: from mail-bw0-f46.google.com ([209.85.214.46]:57884 "EHLO mail-bw0-f46.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752477Ab1BXPV5 (ORCPT ); Thu, 24 Feb 2011 10:21:57 -0500 Received: by bwz15 with SMTP id 15so1106510bwz.19 for ; Thu, 24 Feb 2011 07:21:56 -0800 (PST) Content-Disposition: inline In-Reply-To: Sender: linux-ide-owner@vger.kernel.org List-Id: linux-ide@vger.kernel.org To: Sylvain LEVEQUE Cc: linux-ide@vger.kernel.org, justin@jmicron.com (cc'ing Justin) Hello, On Thu, Feb 24, 2011 at 03:32:50PM +0100, Sylvain LEVEQUE wrote: > I own a Venus T5C mini box based on a JMicron JM393 controller, used > in pure port multiplier mode (no JBOD, no RAID). It is plugged in > eSATA to a QNAP TS-119 NAS which runs a 2.6.37 kernel on a Marvell > Kirkwood 88F6281. > > When I first started copying data on any disk of the external box, I > realized the write speed was low (read speed is OK). Testing with dd > and various block sizes shows that the larger the block size, the > higher the throughput is. On my NAS, I determined the work buffer size > of the cp command to be approximately 8KB. > > With a 8KB blocksize, when writing a 16MB file with dd on one of the > disks, I average at 6,5 MB/s write speed. It is already a great > improvement compared to the 393 KB/s write speed I obtained on a > 2.6.36 kernel, but it is low compared to results I obtain with higher > blocksizes (52,4 MB/s with 4MB blocksize) or using other hardware > (laptop with Intel ICH9M-E/M controller reaches 286 MB/s with 8KB > blocksize using a 2.6.31 kernel). > > When I take one of the disks out and connect it in eSATA with a basic > external case, I reach 37,6 MB/s with 8KB blocksize. This result is > pretty constant whatever the blocksize value I tested (from 1KB to > 16MB, doubling each time). Sounds like a latency related issue. Is NCQ being used? -- tejun