On Sat 8 Aug 2009 05:50, Ben Warren pondered: > Allesandro, > > Alessandro Rubini wrote: > > I finally fixed the defrag code, testing with NFS as well. > > Didn't take performance figures, tough, for lack of time.
Checking out performance with tftp - server is Linux 2.6.27 located in Germany, client (U-Boot running net/next) is located in the USA. Packets fragmented, and arriving out of order. U-Boot image (214268 bytes) tftpblocksize download rate (bytes) (ms) (bytes/second) 512 48,246 4,441 1024 24,482 8,752 1468 ( 1MTU) 17,243 12,426 2048 12,517 17,118 2948 ( 2MTU) 8,912 24,042 4428 ( 3MTU) 6,167 34,744 4096 6,608 32,425 5908 ( 4MTU) 4,840 44,270 7388 ( 5MTU) 4,042 53,010 8192 3,641 58,848 8868 ( 6MTU) 3,425 62,560 10348 ( 7MTU) 2,974 72,047 11828 ( 8MTU) 2,736 78,314 13308 ( 9MTU) 2,508 85,433 14788 (10MTU) 2,281 93,935 16000 2,233 95,955 16268 (11MTU) 2,174 98,559 So, that is 17-seconds (default - on the master), to 2 seconds. Wow - a 87% reduction! Doing the same with a larger image (default kernel + ext2 file system == 12Meg), gets similar results - goes from 928,688 ms / 12,493 bytes/second (yeah, that is 15 minutes with a tftpblocksize == 1468) to 107,626 ms / 107,866 bytes/second (that is under two minutes with a tftpblocksize == 16000)... Again - an 88% reduction in time spent waiting... So - I think this is well worth the effort for those people who want to use tftp outside of a local network. (on a local network - things do not change that drastically - An 18Meg file with default tftpblocksize (1468 bytes) took 5084ms to download, and with a tftpblocksize of 16268, it about half the time - 2625 ms... Thanks to Alessandro for putting it together. Feel free to add my Signed-off (once the docs have been updated explaining what this is all for). -Robin _______________________________________________ U-Boot mailing list U-Boot@lists.denx.de http://lists.denx.de/mailman/listinfo/u-boot
