| Message ID | c999a175-d02b-beda-7418-95a5b0f2d432@cn.fujitsu.com |
|---|---|
| State | Rejected |
| Delegated to: | Cyril Hrubis |
| Headers | show |
| Series | busy_poll* : revise poll_cmp | expand |
On 19.03.2018 11:18, sunlianwen wrote: > set low latency busy poll to 50 per socket will > spent more time than set low latency busy poll > to 0 per socket. that mean the value of res_50 > is bigger than res_0, so the value of poll_comp > always less than zero and test fail. No, if busy poll is enabled, i.e. value set above 0, we should expect performance gain in this test. With what driver and kernel version you have such results? Is it really supported by that driver? Thanks, Alexey
Hi Alexey You are right, I think is wrong, I debug this case again, and find the driver is virtio_net no support busy poll. Thanks your advise. Best Regards, Lianwen Sun. On 03/19/2018 08:25 PM, Alexey Kodanev wrote: > On 19.03.2018 11:18, sunlianwen wrote: >> set low latency busy poll to 50 per socket will >> spent more time than set low latency busy poll >> to 0 per socket. that mean the value of res_50 >> is bigger than res_0, so the value of poll_comp >> always less than zero and test fail. > No, if busy poll is enabled, i.e. value set above 0, we should expect > performance gain in this test. > > With what driver and kernel version you have such results? > Is it really supported by that driver? > > Thanks, > Alexey > > > <html> <head> <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> </head> <body text="#000000" bgcolor="#C0C0C0"> <p>Hi Alexey</p> <p> You are right, I think is wrong, I debug this case again,</p> <p>and find the driver is virtio_net no support busy poll.</p> <p>Thanks your advise.</p> <p>Best Regards,<br> </p> <p>Lianwen Sun.<br> </p> <br> <div class="moz-cite-prefix">On 03/19/2018 08:25 PM, Alexey Kodanev wrote:<br> </div> <blockquote type="cite" cite="mid:fae2ea47-a416-f2b4-ee2b-f022f14a14a0@oracle.com"> <pre wrap="">On 19.03.2018 11:18, sunlianwen wrote: </pre> <blockquote type="cite"> <pre wrap="">set low latency busy poll to 50 per socket will spent more time than set low latency busy poll to 0 per socket. that mean the value of res_50 is bigger than res_0, so the value of poll_comp always less than zero and test fail. </pre> </blockquote> <pre wrap=""> No, if busy poll is enabled, i.e. value set above 0, we should expect performance gain in this test. With what driver and kernel version you have such results? Is it really supported by that driver? Thanks, Alexey </pre> </blockquote> <br> </body> </html>
On 03/19/2018 08:25 PM, Alexey Kodanev wrote: > On 19.03.2018 11:18, sunlianwen wrote: >> set low latency busy poll to 50 per socket will >> spent more time than set low latency busy poll >> to 0 per socket. that mean the value of res_50 >> is bigger than res_0, so the value of poll_comp >> always less than zero and test fail. > No, if busy poll is enabled, i.e. value set above 0, we should expect > performance gain in this test. > > With what driver and kernel version you have such results? > Is it really supported by that driver? the kernel version is 4.16-rc5, the driver is virtio_net. > Thanks, > Alexey > > > Thanks, Lianwen Sun <html> <head> <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> </head> <body text="#000000" bgcolor="#C0C0C0"> <p><br> </p> <br> <div class="moz-cite-prefix">On 03/19/2018 08:25 PM, Alexey Kodanev wrote:<br> </div> <blockquote type="cite" cite="mid:fae2ea47-a416-f2b4-ee2b-f022f14a14a0@oracle.com"> <pre wrap="">On 19.03.2018 11:18, sunlianwen wrote: </pre> <blockquote type="cite"> <pre wrap="">set low latency busy poll to 50 per socket will spent more time than set low latency busy poll to 0 per socket. that mean the value of res_50 is bigger than res_0, so the value of poll_comp always less than zero and test fail. </pre> </blockquote> <pre wrap=""> No, if busy poll is enabled, i.e. value set above 0, we should expect performance gain in this test. With what driver and kernel version you have such results? Is it really supported by that driver? </pre> </blockquote> the kernel version is 4.16-rc5, the driver is virtio_net.<br> <blockquote type="cite" cite="mid:fae2ea47-a416-f2b4-ee2b-f022f14a14a0@oracle.com"> <pre wrap=""> Thanks, Alexey </pre> </blockquote> Thanks,<br> Lianwen Sun<br> </body> </html>
On 03/20/2018 03:18 AM, sunlianwen wrote: > Hi Alexey > > You are right, I think is wrong, I debug this case again, > > and find the driver is virtio_net no support busy poll. There is support in virtio_net... may be the problem in the underlying configuration/driver, latency between guest and the other host? you could also try netperf -H remote_host -t TCP_RR with/without busy_polling: # sysctl net.core.busy_read=50 # sysctl net.core.busy_poll=50 Thanks, Alexey
Hi Alexey On 03/21/2018 06:50 PM, Alexey Kodanev wrote: > On 03/20/2018 03:18 AM, sunlianwen wrote: >> Hi Alexey >> >> You are right, I think is wrong, I debug this case again, >> >> and find the driver is virtio_net no support busy poll. > > There is support in virtio_net... may be the problem in the underlying > configuration/driver, latency between guest and the other host? you could > also try netperf -H remote_host -t TCP_RR with/without busy_polling: > > # sysctl net.core.busy_read=50 > # sysctl net.core.busy_poll=50 > Thanks your advise. and I find a patch:"virtio_net: remove custom busy_poll" patch link :https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/drivers/net/virtio_net.c?id=ceef438d613f6d I am not sure whether this patch mean virtio_net no support busy poll. Below is debuginfo follow your advise. # sysctl net.core.busy_read=0 net.core.busy_read = 0 # sysctl net.core.busy_poll=0 net.core.busy_poll = 0 # netperf -H 192.168.122.248 -t TCP_RR MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 192.168.122.248 () port 0 AF_INET : first burst 0 Local /Remote Socket Size Request Resp. Elapsed Trans. Send Recv Size Size Time Rate bytes Bytes bytes bytes secs. per sec 16384 87380 1 1 10.00 30101.63 16384 87380 # sysctl net.core.busy_read=50 net.core.busy_read = 50 # sysctl net.core.busy_poll=50 net.core.busy_poll = 50 # netperf -H 192.168.122.248 -t TCP_RR MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 192.168.122.248 () port 0 AF_INET : first burst 0 Local /Remote Socket Size Request Resp. Elapsed Trans. Send Recv Size Size Time Rate bytes Bytes bytes bytes secs. per sec 16384 87380 1 1 10.00 37968.90 16384 87380 ----------------------------------------------------------------------- <<<test_output>>> incrementing stop busy_poll01 1 TINFO: Network config (local -- remote): busy_poll01 1 TINFO: eth1 -- eth1 busy_poll01 1 TINFO: 192.168.1.41/24 -- 192.168.1.20/24 busy_poll01 1 TINFO: fd00:1:1:1::1/64 -- fd00:1:1:1::2/64 busy_poll01 1 TINFO: set low latency busy poll to 50 busy_poll01 1 TINFO: run server 'netstress -R 500000 -B /tmp/ltp-EmybkMxKgu/busy_poll01.IIOgfKYQ6P' busy_poll01 1 TINFO: run client 'netstress -l -H 192.168.1.20 -a 2 -r 500000 -d res_50 -g 44175' busy_poll01 1 TPASS: netstress passed, time spent '53265' ms busy_poll01 2 TINFO: set low latency busy poll to 0 busy_poll01 2 TINFO: run server 'netstress -R 500000 -B /tmp/ltp-EmybkMxKgu/busy_poll01.IIOgfKYQ6P' busy_poll01 2 TINFO: run client 'netstress -l -H 192.168.1.20 -a 2 -r 500000 -d res_0 -g 46767' busy_poll01 2 TPASS: netstress passed, time spent '23393' ms busy_poll01 3 TFAIL: busy poll result is '-127' % <<<execution_status>>> initiation_status="ok" duration=79 termination_type=exited termination_id=1 corefile=no cutime=148 cstime=6930 <<<test_end>>> INFO: ltp-pan reported some tests FAIL LTP Version: 20180118 ############################################################### Done executing testcases. LTP Version: 20180118 ############################################################### Thanks, Lianwen Sun <html> <head> <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> </head> <body text="#000000" bgcolor="#C0C0C0"> <p>Hi Alexey<br> </p> <div class="moz-cite-prefix">On 03/21/2018 06:50 PM, Alexey Kodanev wrote:<br> </div> <blockquote type="cite" cite="mid:4f48b8e7-21ca-088e-24eb-c5b79c3bac18@oracle.com"> <pre wrap="">On 03/20/2018 03:18 AM, sunlianwen wrote: </pre> <blockquote type="cite"> <pre wrap="">Hi Alexey You are right, I think is wrong, I debug this case again, and find the driver is virtio_net no support busy poll. </pre> </blockquote> <pre wrap=""> There is support in virtio_net... may be the problem in the underlying configuration/driver, latency between guest and the other host? you could also try netperf -H remote_host -t TCP_RR with/without busy_polling: # sysctl net.core.busy_read=50 # sysctl net.core.busy_poll=50 </pre> </blockquote> Thanks your advise. and I find a patch:"virtio_net: remove custom busy_poll"<br> patch link :<a class="moz-txt-link-freetext" href="https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/drivers/net/virtio_net.c?id=ceef438d613f6d">https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/drivers/net/virtio_net.c?id=ceef438d613f6d</a><br> I am not sure whether this patch mean virtio_net no support busy poll.<br> <br> Below is debuginfo follow your advise.<br> <br> # sysctl net.core.busy_read=0<br> net.core.busy_read = 0<br> <br> # sysctl net.core.busy_poll=0<br> net.core.busy_poll = 0<br> <br> # netperf -H 192.168.122.248 -t TCP_RR<br> MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 192.168.122.248 () port 0 AF_INET : first burst 0<br> Local /Remote<br> Socket Size Request Resp. Elapsed Trans.<br> Send Recv Size Size Time Rate<br> bytes Bytes bytes bytes secs. per sec<br> <br> 16384 87380 1 1 10.00 30101.63<br> 16384 87380<br> <br> # sysctl net.core.busy_read=50<br> net.core.busy_read = 50<br> # sysctl net.core.busy_poll=50<br> net.core.busy_poll = 50<br> <br> # netperf -H 192.168.122.248 -t TCP_RR<br> MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 192.168.122.248 () port 0 AF_INET : first burst 0<br> Local /Remote<br> Socket Size Request Resp. Elapsed Trans.<br> Send Recv Size Size Time Rate<br> bytes Bytes bytes bytes secs. per sec<br> <br> 16384 87380 1 1 10.00 37968.90<br> 16384 87380<br> <br> -----------------------------------------------------------------------<br> <<<test_output>>><br> incrementing stop<br> busy_poll01 1 TINFO: Network config (local -- remote):<br> busy_poll01 1 TINFO: eth1 -- eth1<br> busy_poll01 1 TINFO: 192.168.1.41/24 -- 192.168.1.20/24<br> busy_poll01 1 TINFO: fd00:1:1:1::1/64 -- fd00:1:1:1::2/64<br> busy_poll01 1 TINFO: set low latency busy poll to 50<br> busy_poll01 1 TINFO: run server 'netstress -R 500000 -B /tmp/ltp-EmybkMxKgu/busy_poll01.IIOgfKYQ6P'<br> busy_poll01 1 TINFO: run client 'netstress -l -H 192.168.1.20 -a 2 -r 500000 -d res_50 -g 44175'<br> busy_poll01 1 TPASS: netstress passed, time spent '53265' ms<br> busy_poll01 2 TINFO: set low latency busy poll to 0<br> busy_poll01 2 TINFO: run server 'netstress -R 500000 -B /tmp/ltp-EmybkMxKgu/busy_poll01.IIOgfKYQ6P'<br> busy_poll01 2 TINFO: run client 'netstress -l -H 192.168.1.20 -a 2 -r 500000 -d res_0 -g 46767'<br> busy_poll01 2 TPASS: netstress passed, time spent '23393' ms<br> busy_poll01 3 TFAIL: busy poll result is '-127' %<br> <<<execution_status>>><br> initiation_status="ok"<br> duration=79 termination_type=exited termination_id=1 corefile=no<br> cutime=148 cstime=6930<br> <<<test_end>>><br> INFO: ltp-pan reported some tests FAIL<br> LTP Version: 20180118<br> <br> ###############################################################<br> <br> Done executing testcases.<br> LTP Version: 20180118<br> ###############################################################<br> <br> Thanks,<br> Lianwen Sun<br> </body> </html>
On 22.03.2018 11:03, sunlianwen wrote: > Hi Alexey > > On 03/21/2018 06:50 PM, Alexey Kodanev wrote: >> On 03/20/2018 03:18 AM, sunlianwen wrote: >>> Hi Alexey >>> >>> You are right, I think is wrong, I debug this case again, >>> >>> and find the driver is virtio_net no support busy poll. >> >> There is support in virtio_net... may be the problem in the underlying >> configuration/driver, latency between guest and the other host? you could >> also try netperf -H remote_host -t TCP_RR with/without busy_polling: >> >> # sysctl net.core.busy_read=50 >> # sysctl net.core.busy_poll=50 >> > Thanks your advise. and I find a patch:"virtio_net: remove custom busy_poll" > patch link :https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/drivers/net/virtio_net.c?id=ceef438d613f6d > I am not sure whether this patch mean virtio_net no support busy poll. It means it is using generic busy polling implementation now. > > Below is debuginfo follow your advise. > > # sysctl net.core.busy_read=0 > net.core.busy_read = 0 > > # sysctl net.core.busy_poll=0 > net.core.busy_poll = 0 > > # netperf -H 192.168.122.248 -t TCP_RR > MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 192.168.122.248 () port 0 AF_INET : first burst 0 > Local /Remote > Socket Size Request Resp. Elapsed Trans. > Send Recv Size Size Time Rate > bytes Bytes bytes bytes secs. per sec > > 16384 87380 1 1 10.00 30101.63 > 16384 87380 > > # sysctl net.core.busy_read=50 > net.core.busy_read = 50 > # sysctl net.core.busy_poll=50 > net.core.busy_poll = 50 > > # netperf -H 192.168.122.248 -t TCP_RR > MIGRATED TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 192.168.122.248 () port 0 AF_INET : first burst 0 > Local /Remote > Socket Size Request Resp. Elapsed Trans. > Send Recv Size Size Time Rate > bytes Bytes bytes bytes secs. per sec > > 16384 87380 1 1 10.00 37968.90 > 16384 87380 > Looks like busy polling is working, transaction rate is obviously faster than without busy polling. > ----------------------------------------------------------------------- > <<<test_output>>> > incrementing stop > busy_poll01 1 TINFO: Network config (local -- remote): > busy_poll01 1 TINFO: eth1 -- eth1 > busy_poll01 1 TINFO: 192.168.1.41/24 -- 192.168.1.20/24 You run netperf with 192.168.122.248 server, LTP using the other hosts? Could you check netperf and LTP for the same hosts/interfaces? > busy_poll01 1 TINFO: fd00:1:1:1::1/64 -- fd00:1:1:1::2/6> busy_poll01 1 TINFO: set low latency busy poll to 50 > busy_poll01 1 TINFO: run server 'netstress -R 500000 -B /tmp/ltp-EmybkMxKgu/busy_poll01.IIOgfKYQ6P' > busy_poll01 1 TINFO: run client 'netstress -l -H 192.168.1.20 -a 2 -r 500000 -d res_50 -g 44175' > busy_poll01 1 TPASS: netstress passed, time spent '53265' ms > busy_poll01 2 TINFO: set low latency busy poll to 0 > busy_poll01 2 TINFO: run server 'netstress -R 500000 -B /tmp/ltp-EmybkMxKgu/busy_poll01.IIOgfKYQ6P' > busy_poll01 2 TINFO: run client 'netstress -l -H 192.168.1.20 -a 2 -r 500000 -d res_0 -g 46767' > busy_poll01 2 TPASS: netstress passed, time spent '23393' ms > busy_poll01 3 TFAIL: busy poll result is '-127' % > <<<execution_status>>> > initiation_status="ok" > duration=79 termination_type=exited termination_id=1 corefile=no > cutime=148 cstime=6930 > <<<test_end>>> > INFO: ltp-pan reported some tests FAIL > LTP Version: 20180118 > > ############################################################### > > Done executing testcases. > LTP Version: 20180118 > ############################################################### > > Thanks, > Lianwen Sun
diff --git a/testcases/network/busy_poll/busy_poll01.sh b/testcases/network/busy_poll/busy_poll01.sh index 3c3035600..119ae0176 100755 --- a/testcases/network/busy_poll/busy_poll01.sh +++ b/testcases/network/busy_poll/busy_poll01.sh @@ -60,7 +60,7 @@ for x in 50 0; do tst_netload -H $(tst_ipaddr rhost) -d res_$x done -poll_cmp=$(( 100 - ($(cat res_50) * 100) / $(cat res_0) )) +poll_cmp=$(( 100 - ($(cat res_0) * 100) / $(cat res_50) )) if [ "$poll_cmp" -lt 1 ]; then tst_resm TFAIL "busy poll result is '$poll_cmp' %" diff --git a/testcases/network/busy_poll/busy_poll02.sh b/testcases/network/busy_poll/busy_poll02.sh index 427857996..05b5cf8c4 100755 --- a/testcases/network/busy_poll/busy_poll02.sh +++ b/testcases/network/busy_poll/busy_poll02.sh @@ -51,7 +51,7 @@ for x in 50 0; do tst_netload -H $(tst_ipaddr rhost) -d res_$x -b $x done -poll_cmp=$(( 100 - ($(cat res_50) * 100) / $(cat res_0) )) +poll_cmp=$(( 100 - ($(cat res_0) * 100) / $(cat res_50) )) if [ "$poll_cmp" -lt 1 ]; then tst_resm TFAIL "busy poll result is '$poll_cmp' %"
set low latency busy poll to 50 per socket will spent more time than set low latency busy poll to 0 per socket. that mean the value of res_50 is bigger than res_0, so the value of poll_comp always less than zero and test fail. Signed-off-by: Lianwen Sun <sunlw.fnst@cn.fujitsu.com> --- testcases/network/busy_poll/busy_poll01.sh | 2 +- testcases/network/busy_poll/busy_poll02.sh | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)