[mnet-devel] FEC, blocksizes, chunking

[Jim McCoy]

On Tuesday, April 15, 2003, at 08:18 AM, Hauke Johannknecht wrote:

> Problem: FEC eats up CPU when using too many blocks.

Can someone confirm that this problem also exists in the C code we are 
deriving the fec encoder routines from? One of Rizzo's papers suggests 
that where math is expensive GF(8) and lookup tables should be used, 
but he adds a sectoin in the actual timing notes that suggests that on 
some fairly old boxes he was seeing only a 3-4x slowdown for GF()>8. 
This was expected to balloon to 100x for doing it all on the Java VM 
(which is why the swarmcast-derived Java FEC systems are limited to 
GF(8) iirc) but I am not sure why we should be seeing this slowdown.

> i asserted that we will start to see problems with interrupted
> blocktransfers at such big blocksizes. if this problem arises, it
> can be fixed by adding partial blocktransfers (blocklets). the
> blocklet table could be exchanged between downloader and
> blockserver when trying to get-block, triggered by a user-supplied
> sizelimit.

Any reason not to just use blockid[byte-range slice]? Is there any 
reason for a blocklet to be an independently addressable blob of data?

Jim



-------------------------------------------------------
This sf.net email is sponsored by:ThinkGeek
Welcome to geek heaven.
http://thinkgeek.com/sf
_______________________________________________
mnet-devel mailing list
mnet-devel at lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/mnet-devel