[mnet-devel] fec performance measured, part 2

[Zooko]

Hey waitasec -- the superlinearity isn't due to the scattered reads at all -- 
it's due to the superlinearity of Rizzo.  fectest makes the packet size be 
2 MB, which means that a 128 MB file will have M = 64 and a 256 MB file will 
have M = 128.

I ran the previous test from my tmpfs to confirm that.  It behaved the same 
way from tmpfs as it did from reiserfs, demontrating that disk access is not 
an issue (at these file sizes).

Here are the times to encode a 256 MB with different share sizes:

share size	num shares	time (secs)
----------	----------	-----------
      2 MB	       128		283
      4 MB              64	        142
      8 MB	        32 	         72
     16 MB	        16	         37
     32 MB	         8	         19
     64 MB	         4	         14
    128 MB	         2	         29
    256 MB	         1	         48

Note that, so far as we have tested, encoding is *not* superlinear in 
filesize, only in sharenum, so if it takes 37 seconds to encode a 256 MB file 
with sharenum=16, then it ought to take 74 seconds to encode a 512 MB file 
with sharenum=16, and so forth.

--Zooko

P.S.  Attached is the variant of fectest.py used to generate these stats -- it 
takes sharesize as the first command-line argument.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: text/python
Size: 602 bytes
Desc: not available
Url : http://lists.mnetproject.org/pipermail/mnet-devel/attachments/20030724/43caa882/attachment-0001.bin