fxp performance with POLLING
Bartosz Stec
admin at kkip.pl
Mon Oct 6 09:19:06 UTC 2008
Jeremy Chadwick pisze:
> On Mon, Oct 06, 2008 at 09:02:33AM +0200, Bartosz Stec wrote:
>
>> Alfred Perlstein wrote:
>>
>>> * Bartosz Stec <admin at kkip.pl> [081003 07:23] wrote:
>>>
>>>
>>>> Hello again :)
>>>>
>>>> With POLLING enabled I experience about 10%-25% performance drop when
>>>> copying files over network. Tested with both SAMBA and NFS. Is it
>>>> normal?
>>>>
>>>> FreeBSD 7.1-PRERELEASE #0: Sat Sep 6 01:52:12 CEST 2008
>>>> fxp0: <Intel 82801DB (ICH4) Pro/100 Ethernet> port 0xc800-0xc83f mem
>>>> 0xe1021000-0xe1021fff irq 20 at device 8.0 on pci1
>>>>
>>>> # ifconfig fxp0
>>>> fxp0: flags=9843<UP,BROADCAST,RUNNING,SIMPLEX,LINK0,MULTICAST>
>>>> metric 0 mtu 1500
>>>> options=8<VLAN_MTU>
>>>> ether 00:20:ed:42:87:13
>>>> inet 192.168.0.2 netmask 0xffffff00 broadcast 192.168.0.255
>>>> media: Ethernet autoselect (100baseTX <full-duplex>)
>>>> status: active
>>>>
>>>> BTW overall SAMBA performance still sucks on 7.1-pre as much as on
>>>> RELENG_5 ...:( - 7.5 MB/s peak.
>>>>
>>>>
>>> 7.5MB is 75% effeciency of a 100mbit card. Not amazing, but
>>> not "sucks".
>>>
>>> Where do you see faster performance?
>>>
>>> Between windows machines on the same hardware or linux server?
>>>
>>>
>>>
>> It sucks because it is a peak performance. About 5-6 MB/s average. I
>> tried polling only because I found some suggestions on mailing lists,
>> that it could improve performance with SAMBA on FreeBSD. As you see at
>> the top of this thread - not in my case :) I also tried sysctl tunings,
>> and smb.conf settings, also suggested on maling lists, with no or very
>> little improvements noticed. Most of suggestions unfortunately end with
>> "change OS to Linux if you want to use SAMBA". I think I will try to
>> change NIC to 1Gbit - hope that helps :) Or maybe there's some "FreeBSD
>> and SAMBA tuning guide" which I didn't found?
>>
>
> Can you please test network I/O using something like netperf or one of
> the other network-benchmark tools and not things like NFS or Samba
> which rely on disk I/O and other aspects?
>
>
OK
It was first time i was using nerperf so I'm not sure I did it
correctly. I installed netperf port on SAMBA serwer (IP 192.168.0.2),
and also download windows binary to windows xp machine (IP
192.168.0.10). All tests ran for one minute.
First test - netperf on FreeBSD and netserver on Windows:
# netperf -l 60 -t TCP_STREAM -H 192.168.0.10
TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
192.168.0.10 (192.168.0.10) port 0 AF_INET
Recv Send Send
Socket Socket Message Elapsed
Size Size Size Time Throughput
bytes bytes bytes secs. 10^6bits/sec
8192 32768 32768 60.00 93.97
# netperf -l 60 -t TCP_SENDFILE -H 192.168.0.10
TCP SENDFILE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
192.168.0.10 (192.168.0.10) port 0 AF_INET
Recv Send Send
Socket Socket Message Elapsed
Size Size Size Time Throughput
bytes bytes bytes secs. 10^6bits/sec
8192 32768 32768 60.00 93.45
# netperf -l 60 -t TCP_RR -H 192.168.0.10
TCP REQUEST/RESPONSE TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to
192.168.0.10 (192.168.0.10) port 0 AF_INET
Local /Remote
Socket Size Request Resp. Elapsed Trans.
Send Recv Size Size Time Rate
bytes Bytes bytes bytes secs. per sec
32768 65536 1 1 60.00 2433.99
8192 8192
# ifconfig fxp0
fxp0: flags=9843<UP,BROADCAST,RUNNING,SIMPLEX,LINK0,MULTICAST>
metric 0 mtu 1500
options=8<VLAN_MTU>
ether 00:20:ed:42:87:13
inet 192.168.0.2 netmask 0xffffff00 broadcast 192.168.0.255
media: Ethernet autoselect (100baseTX <full-duplex>)
status: active
Second test - netperf on Windows and netserver on FreeBSD:
Unfortunately won't run:
C:\software>netperf-a4 -l 60 -H 192.168.0.2
TCP STREAM TEST to 192.168.0.2
recv_response: partial response received: 0 bytes
Hovewer, thanks to Alfred Perlstein who send mefollowing link:
http://www.mavetju.org/mail/view_message.php?list=freebsd-net&id=755111&thread=no&tag=yes,
I set SO_SNBUF and SO_RCVBUF in smb.conf to 2920. Without any additional
tuning in sysctl I now got about 8MB/s which is *much* better result
than before. It still could be better than that if I am reading netpertf
results correctly :)
Thanks Alfred!
--
Bartosz Stec
More information about the freebsd-stable
mailing list