Menu
▾
▴
RE: [coLinux-users] NFS performance and colinux
|
From: Claude L. (QB/EMC) <cla...@er...> - 2004-08-20 13:36:48
|
Hi George, I have tested successfully option 1 from the list provided in my = previous message. So, I have been able to make a Solaris jumpstart from = a Windows 2000 based system using coLinux 0.6.2 (snapshot-20040710) ! I = have recompiled the 2.6.7 colinux kernel with the "NFS over TCP" = feature enabled. It works like a charm. The installation speed is very good. It's quite = the same speed than a native Linux system. Of course, the IP = fragmentation/reassembly problem I have observed using the bridged = network configuration under win2k is not solved but at least there is a = workaround. As far as I can see, there is no more = fragmentation/reassembly problem using the new NFS feature because I = think each NFS packet subject to fragmentation is replaced by a "RPC = continuation" packet. If a part of a packet is missing, I guess only = the missing element will re-transmit instead of the whole NFS stream in = a normal world. This is due to the TCP nature and design. So, the "NFS = over TCP" feature improves NFS reliability for connections under "bad" = networking conditions. Thanks, Claude. Claude LeFran=E7ois=20 cla...@er... -----Original Message----- From: Claude LeFrancois (QB/EMC)=20 Sent: Wednesday, August 18, 2004 5:46 PM To: 'gboutwel' Cc: col...@li... Subject: RE: [coLinux-users] NFS performance and colinux Hi George, I have installed snapshot-20040710. I have made the modification to the = xml file to attach to my Broadcom interface. I have booted my Fedora = Core 1 distribution with the 2.6.7 kernel without a problem. This = snapshot does not contain the nfsd support. So, I have recompiled the = colinux kernel with the appropriate settings. I have also learned that = the NFS daemon service has changed in the 2.6 kernels. So, I had to = make a little adaptation (mount -t nfsd nfsd /proc/fs/nfs). Finally, I = got my colinux working fine. :) However, my Solaris jumpstart does not work yet. I hit the same = problem. It seems the IP fragmentation problem is still there but I = have noticed an good improvement. It looks like the jumpstart client = (the Solaris box) sends a bunch of large NFS read packets at the same = time. Colinux is not able to reply to all of them in time and the = jumpstart client sends ICMP packets with TTL exceeded to colinux. At = this moment the jumpstart client gives "NFS server not responding". The = only way I have to improve this situation is to set the block size to = 1024 which avoids the fragmentation. The bad stuff is that only the = root fs benefits from this settings all the other NFS mounts from the = client are 4K blocks. The jumpstart client is point-to-point connected with my laptop using a = cross-over cable. To solve this situation, I see two options: 1. Hard code the NFS packet size to 1024 in the kernel 2. Try the new feature: NFS over TCP But, the fragmentation problem in colinux is not solved... I am a bit = confused now. I don't know what to do at this point. I am using the = bridged networking configuration. I guess it has nothing to do with the = TAP drivers, isn't it ? I am using win2k so there is no way to enable = bridging through the TAP interface because it appears this is only = supported in XP. I absolutely need to use bridging because the = jumpstart client initiates the jumpstart process by a RARP request. = Then, I cannot use NAT because these packets won't be redirected to my = colinux distribution unless a port forwarding mechanism exists. My system is a HP NC8000 laptop. It has 512 MB RAM and 1.5 GHz Pentium = M CPU. The network interface is a Broadcom NetXtreme Gigabit adapter. I = have already disabled TCP checksums and flow control on this interface. Any suggestions are really welcome ! Here is the output of my latest ping test: PING 192.168.0.101 (192.168.0.101) 4096(4124) bytes of data. 4104 bytes from 192.168.0.101: icmp_seq=3D1 ttl=3D255 time=3D2.24 ms 4104 bytes from 192.168.0.101: icmp_seq=3D2 ttl=3D255 time=3D2.10 ms 4104 bytes from 192.168.0.101: icmp_seq=3D3 ttl=3D255 time=3D2.23 ms 4104 bytes from 192.168.0.101: icmp_seq=3D7 ttl=3D255 time=3D2.09 ms 4104 bytes from 192.168.0.101: icmp_seq=3D8 ttl=3D255 time=3D2.22 ms 4104 bytes from 192.168.0.101: icmp_seq=3D10 ttl=3D255 time=3D2.10 ms 4104 bytes from 192.168.0.101: icmp_seq=3D12 ttl=3D255 time=3D2.23 ms 4104 bytes from 192.168.0.101: icmp_seq=3D13 ttl=3D255 time=3D2.22 ms 4104 bytes from 192.168.0.101: icmp_seq=3D15 ttl=3D255 time=3D2.22 ms 4104 bytes from 192.168.0.101: icmp_seq=3D17 ttl=3D255 time=3D2.06 ms 4104 bytes from 192.168.0.101: icmp_seq=3D20 ttl=3D255 time=3D2.23 ms 4104 bytes from 192.168.0.101: icmp_seq=3D23 ttl=3D255 time=3D2.22 ms 4104 bytes from 192.168.0.101: icmp_seq=3D24 ttl=3D255 time=3D2.23 ms 4104 bytes from 192.168.0.101: icmp_seq=3D26 ttl=3D255 time=3D2.09 ms 4104 bytes from 192.168.0.101: icmp_seq=3D27 ttl=3D255 time=3D2.11 ms 4104 bytes from 192.168.0.101: icmp_seq=3D30 ttl=3D255 time=3D2.22 ms 4104 bytes from 192.168.0.101: icmp_seq=3D32 ttl=3D255 time=3D1.97 ms 4104 bytes from 192.168.0.101: icmp_seq=3D33 ttl=3D255 time=3D2.09 ms 4104 bytes from 192.168.0.101: icmp_seq=3D34 ttl=3D255 time=3D2.07 ms --- 192.168.0.101 ping statistics --- 35 packets transmitted, 19 received, 45% packet loss, time 34196ms rtt min/avg/max/mdev =3D 1.974/2.158/2.242/0.087 ms, pipe 2 Thanks, Claude. Claude LeFran=E7ois=20 cla...@er... -----Original Message----- From: col...@li... [mailto:col...@li...]On Behalf Of gboutwel Sent: Wednesday, August 18, 2004 9:32 AM To: col...@li... Cc: Claude LeFrancois (QB/EMC) Subject: Re: [coLinux-users] NFS performance and colinux > > One question: are you using bridged networking ? > > I'm using the tap adapter, bridged by WinXP with my network card. I use both. On the PCAP bridging there is, also, 0% loss of packets using latest sources. > Not sure about that (I made no tests), but it seems the tap-win32 > adapter is somewhat faster and reliable than the bridged adapter. Also, > the code was revised lately to accomodate larger packets transfers. Yeah, most the tests that Dan has done shows TAP to be the faster interface. George ------------------------------------------ Praize? Enter In... http://www.praize.com/ ------------------------------------------------------- SF.Net email is sponsored by Shop4tech.com-Lowest price on Blank Media 100pk Sonic DVD-R 4x for only $29 -100pk Sonic DVD+R for only $33 Save 50% off Retail on Ink & Toner - Free Shipping and Free Gift. http://www.shop4tech.com/z/Inkjet_Cartridges/9_108_r285 _______________________________________________ coLinux-users mailing list coL...@li... https://lists.sourceforge.net/lists/listinfo/colinux-users |