I have a file server (serverworks dual pIII motherboard) with a Syskonnect 9821 gigabit card that I want to be able to transfer files very quickly using ftp and http. It's connected to a 16 port wirespeed gige switch and two other machines. I run the latest Redhat 7.3 supplied SMP kernel (2.4.18-19.7.xsmp). I have made changes to tcp window size, transfer buffers etc according to recommendations. First I wanted to establish a baseline, so I did some transfers using ftp via loopback. One single transfer yields 70-80 megabyte/s, 3 in parallell yields an aggregate of 120-140 megabyte/s. Internally, I get numbers I expect. These are 500 megabyte files transferred from disk cache into /dev/null using wget. Now, over the LAN, I do one transfer from another machine, I get 30-50 megabyte/s. This is from a machine connected to the same switch, also using a 64bit/66MHz Syskonnect card. Now, from the third machine, I start another transfer at the same time as the first. I now get 15-25 meg/s if I add both transfers (6-14 megabyte/s each). I see the same results using http (apache web server). It's noticable slower with two transfers than with one. Doesnt matter if one transfer is ftp and the other http, same problem. Using http I can sometimes reach as high as 50 meg/s over the network, I never get even close using two transfers (they immediately bog down to 5-15 meg/s each). Local loopback transfers are never affected by the "network slowdown effect", two network transfers at aggregate 15 meg/s still enables 70 meg/s loopback transfers. The file server is noticable slower in interactive (remote login) performance when doing two parallell transfers. Pinging over the local lan gives ping times of sometimes as high as 50ms during the transfers. As soon as the transfers are ended I go back to .06 ms ping time. Could this be an interrupt handling problem so I might be better off using "noapic", or even an SMP issue so I would be better off actually only using one cpu (they're both 1.13GHz pIII:s with 512 kb cache). Any other recommendations, I feel I should be able to quite easily reach at least 50-70 megabyte/s aggregate over the network with my setup (as long as disk I/O is not an issue of course, all the above is with file pre-cached on the file server). I actually feel that my old motherboard (VP6 with dual pIII 1GHz processors, all 33Mhz 32bit PCI busses) gave me better results when it came to handling several simultanious requests, whereas the new serverworks motherboard really kicks ass in all other areas (100+ megabyte/s actual read speed off of a RAID1 of SCSI drives). Any input appreciated. -- Mikael Abrahamsson email: swmike@swm.pp.se - : send the line "unsubscribe linux-net" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
