Folks -
I have just checked-in to the top-of-trunk netperf
(http://www.netperf.org/svn/netperf2/trunk) some changes which enable
use of sendto(MSG_FASTOPEN) in a TCP_CRR test. While I was checking the
system calls I noticed that netperf was calling enable_enobufs() for
every migrated test, not just a UDP_STREAM test (the one where it is
needed), so I fixed that at the same time. Baseline is taken with the
fix in place.
MSG_FASTOPEN is used when one adds a test-specific -F option to the
netperf command line:
netperf -t TCP_CRR -H <destination> -- -F
Just testing the client side from a VM on my workstation (running a
net-next pulled just before 16:30 Pacific time) to my workstation itself
I notice the following:
Without MSG_FASTOPEN:
raj@tardy-ubuntu-1204:~/netperf2_trunk$ src/netperf -H tardy.usa.hp.com
-t TCP_CRR -i 3,30
MIGRATED TCP Connect/Request/Response TEST from 0.0.0.0 (0.0.0.0) port 0
AF_INET to tardy.usa.hp.com () port 0 AF_INET : +/-2.500% @ 99% conf. :
demo
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 2092.33
16384 87380
With MSG_FASTOPEN
raj@tardy-ubuntu-1204:~/netperf2_trunk$ src/netperf -H tardy.usa.hp.com
-t TCP_CRR -i 3,30 -- -F
MIGRATED TCP Connect/Request/Response TEST from 0.0.0.0 (0.0.0.0) port 0
AF_INET to tardy.usa.hp.com () port 0 AF_INET : +/-2.500% @ 99% conf. :
demo
!!! WARNING
!!! Desired confidence was not achieved within the specified iterations.
!!! This implies that there was variability in the test environment that
!!! must be investigated before going further.
!!! Confidence intervals: Throughput : 6.565%
!!! Local CPU util : 0.000%
!!! Remote CPU util : 0.000%
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 2590.18
16384 87380
There was a ~25% increase in TCP_CRR performance, even without the
server actually accepting the magic TCP option. Is that actually
expected? Admittedly, it eliminates a connect() call, but the sequence
before was:
socket()
getsockopt(SO_SNDBUF)
getsockopt(SO_RCVBUF)
setsockopt(SO_REUSEADDR)
bind()
connect()
sendto(4, "n", 1, 0, NULL, 0) = 1
recvfrom(4, "n", 1, 0, NULL, NULL) = 1
recvfrom(4, "", 1, 0, NULL, NULL) = 0
close(4)
and with MSG_FASTOPEN only the connect() goes away. Unless connect() is
really heavy compared to what sendto(MSG_FASTOPEN) does or something
else has changed wrt behaviour.
Anyway, feel free to kick the tires on the netperf changes and let me
know of any problems you encounter.
happy benchmarking,
rick jones
Hi Rick,
On Mon, Aug 6, 2012 at 6:47 PM, Rick Jones [off-list ref] wrote:
Folks -
I have just checked-in to the top-of-trunk netperf
(http://www.netperf.org/svn/netperf2/trunk) some changes which enable use of
sendto(MSG_FASTOPEN) in a TCP_CRR test. While I was checking the system
calls I noticed that netperf was calling enable_enobufs() for every migrated
test, not just a UDP_STREAM test (the one where it is needed), so I fixed
that at the same time. Baseline is taken with the fix in place.
MSG_FASTOPEN is used when one adds a test-specific -F option to the netperf
command line:
netperf -t TCP_CRR -H <destination> -- -F
Just testing the client side from a VM on my workstation (running a net-next
pulled just before 16:30 Pacific time) to my workstation itself I notice the
following:
Without MSG_FASTOPEN:
raj@tardy-ubuntu-1204:~/netperf2_trunk$ src/netperf -H tardy.usa.hp.com -t
TCP_CRR -i 3,30
MIGRATED TCP Connect/Request/Response TEST from 0.0.0.0 (0.0.0.0) port 0
AF_INET to tardy.usa.hp.com () port 0 AF_INET : +/-2.500% @ 99% conf. :
demo
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 2092.33
16384 87380
With MSG_FASTOPEN
raj@tardy-ubuntu-1204:~/netperf2_trunk$ src/netperf -H tardy.usa.hp.com -t
TCP_CRR -i 3,30 -- -F
MIGRATED TCP Connect/Request/Response TEST from 0.0.0.0 (0.0.0.0) port 0
AF_INET to tardy.usa.hp.com () port 0 AF_INET : +/-2.500% @ 99% conf. :
demo
!!! WARNING
!!! Desired confidence was not achieved within the specified iterations.
!!! This implies that there was variability in the test environment that
!!! must be investigated before going further.
!!! Confidence intervals: Throughput : 6.565%
!!! Local CPU util : 0.000%
!!! Remote CPU util : 0.000%
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 2590.18
16384 87380
There was a ~25% increase in TCP_CRR performance, even without the server
actually accepting the magic TCP option. Is that actually expected?
I have a locally enhanced netperf for TCP_CRR over Fast Open and I've
noticed the numbers can change drastically between runs. I have not got
the time to investigate why. (Does it have to do with scheduler and CPU
locality?) How consistent is your perf number?
I plan to submit the server side code soon and will work with you to add
the server side support to TCP_CRR (it requires a new TCP_FASTOPEN
socket option to enable Fast Open on a listener.)
Thanks,
Jerry
Admittedly, it eliminates a connect() call, but the sequence before was:
socket()
getsockopt(SO_SNDBUF)
getsockopt(SO_RCVBUF)
setsockopt(SO_REUSEADDR)
bind()
connect()
sendto(4, "n", 1, 0, NULL, 0) = 1
recvfrom(4, "n", 1, 0, NULL, NULL) = 1
recvfrom(4, "", 1, 0, NULL, NULL) = 0
close(4)
and with MSG_FASTOPEN only the connect() goes away. Unless connect() is
really heavy compared to what sendto(MSG_FASTOPEN) does or something else
has changed wrt behaviour.
Anyway, feel free to kick the tires on the netperf changes and let me know
of any problems you encounter.
happy benchmarking,
rick jones
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
quoted
There was a ~25% increase in TCP_CRR performance, even without the server
actually accepting the magic TCP option. Is that actually expected?
I have a locally enhanced netperf for TCP_CRR over Fast Open and I've
noticed the numbers can change drastically between runs. I have not got
the time to investigate why. (Does it have to do with scheduler and CPU
locality?) How consistent is your perf number?
I recall the performance being reasonably consistent but there has been
a vacation of my own in the middle there so my dimm memory is a bit fuzzy :)
Historically (and going beyond just Linux) a TCP_CRR test can have some
non-trivial run to run variation thanks to (attempted) TIME_WAIT reuse.
Whether that is happening to you I don't know, but it might be worth a
look.
I plan to submit the server side code soon and will work with you to add
the server side support to TCP_CRR (it requires a new TCP_FASTOPEN
socket option to enable Fast Open on a listener.)
Works for me.
happy benchmarking,
rick jones