Suzanne McIntosh
2007-Mar-26 19:47 UTC
[Xen-devel] Seeking performance numbers for flip vs. copy
Hi, I am looking for performance numbers comparing the network driver flip vs. copy feature on a 32-bit machine. Could not locate this info in Xen archives. Would appreciate any help. Thanks, Sue _______________________________________________________________ Suzanne McIntosh IBM T.J. Watson Research Center Secure Systems and Services _______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel
Santos, Jose Renato G
2007-Mar-26 22:25 UTC
RE: [Xen-devel] Seeking performance numbers for flip vs. copy
From: xen-devel-bounces@lists.xensource.com [mailto:xen-devel-bounces@lists.xensource.com] On Behalf Of Suzanne McIntosh Sent: Monday, March 26, 2007 12:48 PM To: xen-devel@lists.xensource.com Cc: Suzanne McIntosh Subject: [Xen-devel] Seeking performance numbers for flip vs. copy Hi, I am looking for performance numbers comparing the network driver flip vs. copy feature on a 32-bit machine. Could not locate this info in Xen archives. Would appreciate any help. Thanks, Sue _______________________________________________________________ Suzanne McIntosh IBM T.J. Watson Research Center Secure Systems and Services Here is some old data that I have for comparing page flip with copy. Note that this for Xen unstable as of October 3, 2006. I don''t more recent data, but I believe the relative performance should not have changed much, although the absolute numbers may be different now. I hope this helps Renato> Here are the results comparing copy and page flip. > These results were using netperf TCP_STREAM test which use large size > packets. > I run experiments for receiving and transmiting to from/to a external > client. > The results show that copying use less cycles than page flip but dom0 > has higher utilization (due to the copy). > This caused dom0 cpu to saturate and reduced the throuhgput slightlyin> my machine. I have a 4way Intel Xeon 2.8Ghz with 4GB of RAM (eachdomain> is using 512 MB of RAM). > Copy is also better for the transmit experiment since the overhead for> processing ACKs is reduced. > > I will work on the transmit side optimizations starting tomorrow. > > Regards > > Renato > > ======================================================> receive: > ------------------------------------------------------_ > | | | cycles/packet | > | Rate | CPU utilization | (thousands) | > | Mb/s | dom0 dom1 tot | dom0 dom1 tot | > ------------------------------------------------------- > flip | 941.34 | 84.2% 76.3% 160.4% | 30.1 27.2 57.3 | > copy | 907.71 | 99.7% 45.2% 144.9% | 36.9 16.7 53.6 | > ------------------------------------------------------- > > transmit: > ------------------------------------ > | | | > | Rate | CPU utilization | > | Mb/s | dom0 dom1 tot | > ------------------------------------ > flip | 941.31 | 49.6% 43.9% 93.5% | > copy | 941.30 | 45.2% 34.5% 79.7% | > ------------------------------------________________________________ _______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel
Suzanne McIntosh
2007-Mar-27 18:14 UTC
RE: [Xen-devel] Seeking performance numbers for flip vs. copy
Thanks so much - this is very helpful. For comparison purposes, it would be great to know the message size. Also wondering if doms were pinned? Regards, Sue _______________________________________________________________ Suzanne McIntosh IBM T.J. Watson Research Center Secure Systems and Services "Santos, Jose Renato G" <joserenato.santos@hp.com> Sent by: xen-devel-bounces@lists.xensource.com 03/26/2007 06:25 PM To Suzanne McIntosh/Watson/IBM@IBMUS cc Xen Development Mailing List <xen-devel@lists.xensource.com> Subject RE: [Xen-devel] Seeking performance numbers for flip vs. copy From: xen-devel-bounces@lists.xensource.com [mailto:xen-devel-bounces@lists.xensource.com] On Behalf Of Suzanne McIntosh Sent: Monday, March 26, 2007 12:48 PM To: xen-devel@lists.xensource.com Cc: Suzanne McIntosh Subject: [Xen-devel] Seeking performance numbers for flip vs. copy Hi, I am looking for performance numbers comparing the network driver flip vs. copy feature on a 32-bit machine. Could not locate this info in Xen archives. Would appreciate any help. Thanks, Sue _______________________________________________________________ Suzanne McIntosh IBM T.J. Watson Research Center Secure Systems and Services Here is some old data that I have for comparing page flip with copy. Note that this for Xen unstable as of October 3, 2006. I don''t more recent data, but I believe the relative performance should not have changed much, although the absolute numbers may be different now. I hope this helps Renato> Here are the results comparing copy and page flip. > These results were using netperf TCP_STREAM test which use large size > packets. > I run experiments for receiving and transmiting to from/to a external > client. > The results show that copying use less cycles than page flip but dom0 > has higher utilization (due to the copy). > This caused dom0 cpu to saturate and reduced the throuhgput slightlyin> my machine. I have a 4way Intel Xeon 2.8Ghz with 4GB of RAM (eachdomain> is using 512 MB of RAM). > Copy is also better for the transmit experiment since the overhead for> processing ACKs is reduced. > > I will work on the transmit side optimizations starting tomorrow. > > Regards > > Renato > > ======================================================> receive: > ------------------------------------------------------_ > | | | cycles/packet | > | Rate | CPU utilization | (thousands) | > | Mb/s | dom0 dom1 tot | dom0 dom1 tot | > ------------------------------------------------------- > flip | 941.34 | 84.2% 76.3% 160.4% | 30.1 27.2 57.3 | > copy | 907.71 | 99.7% 45.2% 144.9% | 36.9 16.7 53.6 | > ------------------------------------------------------- > > transmit: > ------------------------------------ > | | | > | Rate | CPU utilization | > | Mb/s | dom0 dom1 tot | > ------------------------------------ > flip | 941.31 | 49.6% 43.9% 93.5% | > copy | 941.30 | 45.2% 34.5% 79.7% | > ------------------------------------________________________________ _______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel _______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel
Santos, Jose Renato G
2007-Mar-27 18:28 UTC
RE: [Xen-devel] Seeking performance numbers for flip vs. copy
This data was for transferring 1500 byte data packets with TCP. Note that the difference of performance on transmit is due to the receive of ACK packets (small size). The domains were not explicitly pinned, but since there were only two domains (dom0 and domU) in a 4-cpu machine they were running on different CPUs. Renato ________________________________ From: xen-devel-bounces@lists.xensource.com [mailto:xen-devel-bounces@lists.xensource.com] On Behalf Of Suzanne McIntosh Sent: Tuesday, March 27, 2007 11:15 AM To: Santos, Jose Renato G Cc: xen-devel@lists.xensource.com; Suzanne McIntosh Subject: RE: [Xen-devel] Seeking performance numbers for flip vs. copy Thanks so much - this is very helpful. For comparison purposes, it would be great to know the message size. Also wondering if doms were pinned? Regards, Sue _______________________________________________________________ Suzanne McIntosh IBM T.J. Watson Research Center Secure Systems and Services "Santos, Jose Renato G" <joserenato.santos@hp.com> Sent by: xen-devel-bounces@lists.xensource.com 03/26/2007 06:25 PM To Suzanne McIntosh/Watson/IBM@IBMUS cc Xen Development Mailing List <xen-devel@lists.xensource.com> Subject RE: [Xen-devel] Seeking performance numbers for flip vs. copy From: xen-devel-bounces@lists.xensource.com [mailto:xen-devel-bounces@lists.xensource.com] On Behalf Of Suzanne McIntosh Sent: Monday, March 26, 2007 12:48 PM To: xen-devel@lists.xensource.com Cc: Suzanne McIntosh Subject: [Xen-devel] Seeking performance numbers for flip vs. copy Hi, I am looking for performance numbers comparing the network driver flip vs. copy feature on a 32-bit machine. Could not locate this info in Xen archives. Would appreciate any help. Thanks, Sue _______________________________________________________________ Suzanne McIntosh IBM T.J. Watson Research Center Secure Systems and Services Here is some old data that I have for comparing page flip with copy. Note that this for Xen unstable as of October 3, 2006. I don''t more recent data, but I believe the relative performance should not have changed much, although the absolute numbers may be different now. I hope this helps Renato > Here are the results comparing copy and page flip. > These results were using netperf TCP_STREAM test which use large size > packets. > I run experiments for receiving and transmiting to from/to a external > client. > The results show that copying use less cycles than page flip but dom0 > has higher utilization (due to the copy). > This caused dom0 cpu to saturate and reduced the throuhgput slightly in > my machine. I have a 4way Intel Xeon 2.8Ghz with 4GB of RAM (each domain > is using 512 MB of RAM). > Copy is also better for the transmit experiment since the overhead for > processing ACKs is reduced. > > I will work on the transmit side optimizations starting tomorrow. > > Regards > > Renato > > ====================================================== > receive: > ------------------------------------------------------_ > | | | cycles/packet | > | Rate | CPU utilization | (thousands) | > | Mb/s | dom0 dom1 tot | dom0 dom1 tot | > ------------------------------------------------------- > flip | 941.34 | 84.2% 76.3% 160.4% | 30.1 27.2 57.3 | > copy | 907.71 | 99.7% 45.2% 144.9% | 36.9 16.7 53.6 | > ------------------------------------------------------- > > transmit: > ------------------------------------ > | | | > | Rate | CPU utilization | > | Mb/s | dom0 dom1 tot | > ------------------------------------ > flip | 941.31 | 49.6% 43.9% 93.5% | > copy | 941.30 | 45.2% 34.5% 79.7% | > ------------------------------------ ________________________________ _______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel _______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel