From mboxrd@z Thu Jan 1 00:00:00 1970 Message-ID: <3BF4A2D4.5080302@embeddededge.com> Date: Fri, 16 Nov 2001 00:23:32 -0500 From: Dan Malek MIME-Version: 1.0 To: Kenneth Johansson Cc: Armin Kuster , "linuxppc-embedded@lists.linuxppc.org" Subject: Re: linuxppc_2_4_devel on walnut References: <3BDD855B.8D3FC72F@inn.ericsson.se> <3BE971BC.34591FF1@inn.ericsson.se> <3BEBF4E5.B60BD9E3@inn.ericsson.se> <3BF2A361.5FBB0246@mvista.com> <3BF4041F.A732B099@inn.ericsson.se> Content-Type: text/plain; charset=us-ascii; format=flowed Sender: owner-linuxppc-embedded@lists.linuxppc.org List-Id: Kenneth Johansson wrote: > But "time dd if=/dev/zero of=testing bs=1024k count=100" on a nfs mouted fs > used to be 18s and is now 21s, a 16% slowdown not that I care much but still > quite a large difference. What versions of kernels/drivers are you comparing? Do you do this back-to-back with the same NFS server, or have you introduced more variables into your testing? This is why specially prepared network performance testers are better suited for discussion. Many versions of the 405GP Ethernet driver didn't use (and didn't have) cache coherency functions properly. It was basically lucky, and when errors occurred you usually didn't see them because the packets were discarded and retransmitted. Recent versions of the driver now use properly mapped and managed cache coherent memory. When I was making the driver changes for the new cache attributes, I was thinking we should be experimenting with cached pages and invalidates/flushes traded off against uncached pages. There will certainly be a performance difference. -- Dan ** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/