Menu
▾
▴
RE: [coLinux-users] NFS performance and colinux
|
From: Claude L. (QB/EMC) <cla...@er...> - 2004-08-17 15:24:42
|
Thanks for the info, Nuno. I will try a more recent build asap. One question: are you using bridged networking ? The IP fragmentation problem was produced under the bridged networking = configuration under Win2K. Regards, Claude. Claude LeFran=E7ois=20 cla...@er... -----Original Message----- From: Nuno Lucas [mailto:lu...@nl...] Sent: Monday, August 16, 2004 7:44 PM To: Claude LeFrancois (QB/EMC) Cc: col...@li... Subject: Re: [coLinux-users] NFS performance and colinux 192.168.1.1 is my router in the internal network, not the host OS. colinux root # ping -s 4096 192.168.1.1 PING 192.168.1.1 (192.168.1.1) 4096(4124) bytes of data. 4104 bytes from 192.168.1.1: icmp_seq=3D1 ttl=3D64 time=3D4.50 ms 4104 bytes from 192.168.1.1: icmp_seq=3D2 ttl=3D64 time=3D3.53 ms 4104 bytes from 192.168.1.1: icmp_seq=3D3 ttl=3D64 time=3D3.48 ms 4104 bytes from 192.168.1.1: icmp_seq=3D4 ttl=3D64 time=3D4.11 ms 4104 bytes from 192.168.1.1: icmp_seq=3D5 ttl=3D64 time=3D3.61 ms 4104 bytes from 192.168.1.1: icmp_seq=3D6 ttl=3D64 time=3D3.33 ms 4104 bytes from 192.168.1.1: icmp_seq=3D7 ttl=3D64 time=3D3.72 ms 4104 bytes from 192.168.1.1: icmp_seq=3D8 ttl=3D64 time=3D3.44 ms 4104 bytes from 192.168.1.1: icmp_seq=3D9 ttl=3D64 time=3D3.36 ms 4104 bytes from 192.168.1.1: icmp_seq=3D10 ttl=3D64 time=3D3.63 ms [....] --- 192.168.1.1 ping statistics --- 120 packets transmitted, 120 received, 0% packet loss, time 120280ms rtt min/avg/max/mdev =3D 3.280/3.761/9.236/0.951 ms And a really remote host: colinux root # ping -s 4096 www.colinux.org PING www.colinux.org (66.35.250.210) 4096(4124) bytes of data. 4104 bytes from vhost.sourceforge.net (66.35.250.210): icmp_seq=3D1 = ttl=3D45=20 time=3D599 ms 4104 bytes from vhost.sourceforge.net (66.35.250.210): icmp_seq=3D2 = ttl=3D45=20 time=3D595 ms 4104 bytes from vhost.sourceforge.net (66.35.250.210): icmp_seq=3D3 = ttl=3D45=20 time=3D595 ms 4104 bytes from vhost.sourceforge.net (66.35.250.210): icmp_seq=3D4 = ttl=3D45=20 time=3D594 ms 4104 bytes from vhost.sourceforge.net (66.35.250.210): icmp_seq=3D5 = ttl=3D45=20 time=3D596 ms 4104 bytes from vhost.sourceforge.net (66.35.250.210): icmp_seq=3D6 = ttl=3D45=20 time=3D595 ms 4104 bytes from vhost.sourceforge.net (66.35.250.210): icmp_seq=3D7 = ttl=3D45=20 time=3D595 ms 4104 bytes from vhost.sourceforge.net (66.35.250.210): icmp_seq=3D8 = ttl=3D45=20 time=3D595 ms 4104 bytes from vhost.sourceforge.net (66.35.250.210): icmp_seq=3D9 = ttl=3D45=20 time=3D625 ms 4104 bytes from vhost.sourceforge.net (66.35.250.210): icmp_seq=3D10=20 ttl=3D45 time=3D597 ms [....] --- www.colinux.org ping statistics --- 31 packets transmitted, 30 received, 3% packet loss, time 30386ms rtt min/avg/max/mdev =3D 593.580/598.933/654.012/11.610 ms I used a special build using the latest source, on a 2.6.8.1 kernel. colinux root # uname -a Linux colinux 2.6.8.1-co-0.6.2 #1 Sun Aug 15 04:33:48 WEST 2004 i686=20 Intel(R) Pentium(R) 4 CPU 1.70GHz GenuineIntel GNU/Linux Regards, ~Nuno Lucas Claude LeFrancois (QB/EMC), dando pulos de alegria, escreveu : > Hi George, >=20 > Thanks for your reply. I will give a try to the new release as soon = as possible. >=20 > However, I have made some NFS troubleshooting and I have finally been = able to go a bit further with the jumpstart and I have been able to = narrow down the problem. The NFS performance problem seems to be = produced by the IP fragmentation. >=20 > The default NFS block size on Linux is set to 4096 bytes (4K). When I = set the rootopts value to 1024, the jumpstart process goes a lot = better. This parameter set the rsize (read block size) option of the = NFS mount request of the root filesystem coming from the jumpstart = client. Unfortunately, all the other NFS mounted filesystems are not = using this parameter and suffer from the fragmentation problem. So, the = jumpstart fails later. >=20 > Basically, there is no fragmentation at 1024 but, at 4096, the = fragmentation is very important. To verify that the fragmentation is = responsible of the bad NFS performance, I have tested ping of 4K to a = remote system from colinux: >=20 > PING 142.133.81.2 (142.133.81.2) 4096(4124) bytes of data. > 4104 bytes from 142.133.81.2: icmp_seq=3D9 ttl=3D255 time=3D82.4 ms > 4104 bytes from 142.133.81.2: icmp_seq=3D14 ttl=3D255 time=3D4.17 ms > 4104 bytes from 142.133.81.2: icmp_seq=3D22 ttl=3D255 time=3D4.04 ms > 4104 bytes from 142.133.81.2: icmp_seq=3D24 ttl=3D255 time=3D5.05 ms > 4104 bytes from 142.133.81.2: icmp_seq=3D25 ttl=3D255 time=3D4.15 ms > 4104 bytes from 142.133.81.2: icmp_seq=3D32 ttl=3D255 time=3D21.0 ms > 4104 bytes from 142.133.81.2: icmp_seq=3D34 ttl=3D255 time=3D4.16 ms > 4104 bytes from 142.133.81.2: icmp_seq=3D78 ttl=3D255 time=3D4.17 ms >=20 > --- 142.133.81.2 ping statistics --- > 121 packets transmitted, 8 received, 93% packet loss, time 120805ms > rtt min/avg/max/mdev =3D 4.048/16.165/82.463/25.656 ms, pipe 2 >=20 > The ping test shows the 93% of packet drops... It looks like the = colinux network adapter suffers from an IP fragmentation/reassembly = performance problem. >=20 > Please, can you try the same ping on a later colinux version (ping -s = 4096 <remote_ip_address>) if you have time and if you can ? So, we = could already know if this problem is addressed with the new releases. >=20 > Thanks, >=20 > Claude LeFran=E7ois=20 > cla...@er... |
