Re: Deadlock problems in binaryswap spu?
Joel Welling <[email protected]>
| Newsgroups | gmane.comp.graphics.chromium.user |
|---|---|
| Message-ID | <[email protected]> |
> Joel Welling wrote:
> > Brian Paul wrote:
> >
> >>Hi Joel,
> >>
> >>If each compute node is going to talk to just one binaryswap node, why
> >>use tilesort? I guess I don't see the need for tilesort at all in
> >>this config.
> >>
> >
> > Well, the idea was to be able to dial up and down the number of tilesort
> > servers as I experimented with the configuration, but as you say it
> > wasn't working because the barriers had been broadcast to all servers.
> > I had tried some experiments using just a 'pack' spu, but apparently
> > not enough of them...
> >
> >
> >>As for the error, if your parallel app is creating a barrier of size N
> >>(N=4?), the tilesort SPU will broadcast the barrier to all the
> >>crservers. If the crserver has less than N clients, it'll deadlock.
> >>
> >>The tilesort SPU broadcasts barrier commands because that's what you'd
> >>want to do if you were doing parallel tilesort (M tilesort nodes all
> >>talking to N rendering nodes).
> >>
> >>Barriers, however, become tricky to use in a multi-tiered
> >>configuration (ex: 3 tiers of nodes A->B->C).
> >>
> >>One thing you might try is the crserver option to disable/ignore the
> >>parallel API (barrier/semaphore) calls:
> >>
> >> servernode.Conf('ignore_papi', 1)
> >>
> >
> > That does indeed do the trick, in combination with use of a 'pack' spu
> > rather than tilesort! I'd wondered what that option was good for; I guess
> > now I know!
>
> Next, could you remove the pack SPU? That is, just put the binary
> swap SPUs on the application nodes. That'd save a network hop. I'm
> assuming you have suitable graphics hardware on the app nodes.
>
Ah, that's the rub- the app nodes have no graphics hardware. These are supercomputer nodes, and they're being assigned 'helper' nodes with graphics hardware on the fly. I was hoping that using this configuration would allow me a larger 'fan-in' between app nodes and rendering nodes, but so far that doesn't seem to be the case.
-Joel
-------------------------------------------------------
This SF.Net email is sponsored by xPML, a groundbreaking scripting language
that extends applications into web and mobile media. Attend the live webcast
and join the prime developer group breaking into this new coding territory!
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=110944&bid=241720&dat=121642