🚀 go-pugleaf

My original article:

|> I would like to know some "throughput" values of the transputer-host
|> link.
|> 
|> When I use iserver, writing 1k blocks to the host gives me a performance
|> of about 20k per second.  Using smaller blocks results in very reduced
|> performance.
|> 
|> I wrote a small program to measure only transfer speed by reading data
|> directly from the link.  The maximum throughput was about 110k per
|> second.  I think the difference results in iserver being a single
|> threaded server, so it can't receive new packets while it processes
|> packets, which reduces performance a lot.

My machine is a 386-20 without cache, I use a B004 compatible board without
interrrupts.  The machine has about 4500 dhrystones.

From: Bob Green <bob@inmos.co.uk>

> Even if iserver were multi-threaded, that wouldn't help, since when the iserver
> protocol was invented, it was specified as being synchronous. i.e. an application
> won't even attempt to submit a second request until a reply to its first request
> has been received.

True, a asynchronous protocol would be needed for a multi-threaded
server.  Of course that requires more complicated software at the
transputer side, too.

From: Charles Francois  <Charles.Francois@sophia.inria.fr>

> ...                         We had 6 transputers for that purpose,
> that were bound to a Sun3/140 via VME 8/32 (B014 from INMOS). Is it
> the one that you are using?  I modified the iserver from INMOS to
> make it work over the Ethernet, so that the transputer network
> would be accessible from a Sparcstation for example.  There was no
> other procress running on the Sun3/140, apart from the usual
> deamons.  Well, the best I could get was about 40Kbytes/sec.
> ...                              I think that there is no miracle
> to be expected from this board, and I was told that it is a general
> complaint.  For fast transfer, maybe you should consider the B016
> (which as 4M of shared memory) or the B300 (direct interface to
> Ethernet).

As it seems, iserver isn't best choice if you have to transfer big
amounts of data between host and transputer.  The only easy tuning which
could be done is modifing the iserver protocol to allow bigger packets,
because my benchmarks showed that there were huge differences between
e.g. 512 and 1024 user data packets.  Of course that doesn't help if
the problem forbids blocking.

For high performance, either i/o devices should be directly connected to
a transputerlink or you use at least your own problem specific protocol.

Michael