[OpenAFS] 1.4.8, Rx Performance Improvements, and a Small Business Innovative Research grant

[Jeffrey Altman]

Chas Williams (CONTRACTOR) wrote:
> In message <20081007.100150.249414631.haba@habanero.pdc.kth.se>,Harald Barth wr
> ites:
>> As xplot was not available as portage I tried xplot->gnuplot.
>> Don't do that. The xplot graphics shows much better what is
>> going on than the converted gnuplot. I used xplot 0.90.7.1
>>from http://www.xplot.org/.
> 
> i added some sample plots from some early results. 
> 
> write.xplot shows a 1m write from a client to the server.  it consists of
> two calls (and i think the retransmit message from rxtrace2xplot is bogus
> here since it doesnt understand the backward direction of the exchange).
> 
> compare this with read.xplot from the same directory.  this is a ready
> of a 1m file with a cache chunksize of 14.  (i think i am missing some
> arrow heads).  note that there is an rx call for chunksize.  this makes
> the plot somewhat difficult to read but you can see that this doesnt
> perform well (perhaps i should put arrows pointing down for one direction
> and point up for the other direction).  with 11 rx datagrams per read
> (2^14/RX_DATAGRAM_SIZE) you can see that rx never really gets out of
> slow start.  you get 3 packets, 4 and then the last 4.  next call, you
> restart at 3 packets again.  the rx window never gets close to filling.
> also, every 11 packets, you need to wait for a another complete roundtrip
> to the server.
> 
> so while rx could be improved, afs' usage thereof could be improved a
> bit.

why is slow start "per call"?  That can't be right.