On Fri 24 Jul 2009 04:04, Alessandro Rubini pondered:
This patch add a quick and dirty defrag step in IP reception. This allows to increase the TFTP block size and get more performance in slow links (but at that point it should be made configurable).
The overhead is negligible, verified with an ARM9 CPU and 10MB data file, changing the server MTU from 1500 to 800 and then 550. However, on a LAN connection, I didn't see advantes with using a 4k block size with default MTU.
I do...
Filesize : 4613551 bytes (4.3Mbytes)
tftp modified, so it doesn't print out hashes (have have things slow down for the UART printing) - to make sure we are getting a true network measurement...
+#ifndef CONFIG_TFTP_QUIET if (((TftpBlock - 1) % 10) == 0) { putc ('#'); } else if ((TftpBlock % (10 * HASHES_PER_LINE)) == 0) puts ("\n\t "); +#endif
block size seconds 512 9.43 144.78% (+2.92 seconds) 1468 6.52 100.00% (default) 2048 6.27 96.20% (-0.25 seconds) 3072 5.69 87.40% (-0.82 seconds) 4096 5.46 83.81% (-1.05 seconds) 5120 5.26 80.76% (-1.25 seconds) 6144 5.13 78.79% (-1.38 seconds) 7168 5.09 78.13% (-1.42 seconds) 8192 5.01 76.92% (-1.50 seconds) 10240 4.94 75.83% (-1.58 seconds) 12288 4.87 74.74% (-1.65 seconds) 14336 4.85 74.49% (-1.66 seconds) 16384 4.82 73.95% (-1.70 seconds)
Saving that 1.7 seconds is worth it to me...
On a slower connection - I would guess that the difference is going to be more dramatic...
I needed to modify your patch a little bit to get it working on my platform.
If Ben/Wolfgang are interested in taking this - I'll fix up my mods, and send it back.
-Robin