Hqueue - Incredibly slow IFD generation on Win7

[danw]

Hey all,

I'll check in with SESI about this after the weekend, but thought I'd see if anyone on here has any pearls of wisdom first.

I'm trying to get Hqueue set up for the first time, and after battling with Windows' firewall settings and the hqueue services' account credentials for a day, I've finally got it up and running correctly!

 

The only problem now is, the IFD generation stage seems to be monumentally slow!  My tester scene just has a basic torus and a camera, nothing else, not even lighting.  I've got the hqueue renderer set to defaults, other than playing around with batching number and CPU count per batch.

 

For one thing, it seems like the Hqueue server is a tad sluggish actually allocating anything... it takes a few seconds at least to allocate each individual task to a client.

 

The worse aspect is that it seems to take the IFD-generating client anything from 5 to 30 seconds to get around to generating a single batch of ifds.  The actual IFD generation appears to be essentially instant, as it should be with such a simple scene, as once it actually gets going on a batch, it doesn't matter if it's a batch of 1 frame or of 50... it'll output instantly.  It just takes ages to start each one.

 

With batched frames set to 1, it's taking a full 30 minutes just to generate 300 frames of ifds!

 

 

I've taken Windows Firewall out of the equation, as I've tested briefly with it disabled completely on the server and the clients.  I've tried having it run Houdini from the network share and from the local installation directory, and that doesn't change anything either.

 

Any ideas what other sort of things should I check?  I get the feeling it's hython that's being sluggish for some reason, but I'm not sure what to do about it.

[danw]

Hmm, on further testing, it does seem like anything involving hython takes ages.  If I submit a job with IFD generation turned off - so that it just creates a "Prepare Render Jobs" hython task, and each batch generates its own ifds using hython - it still manages to take 5 minutes to process the Prepare Render Jobs task, and each batch takes around twice as long as it does when running straight through mantra.  Overall it's a bit quicker I suppose, as it's distributing the ifd generation rather than just waiting on a single job, but it's still slower than just kicking the whole thing off locally on a single workstation.

 

Is hython just innately slow?  Should I be expecting a very simple render task to end up taking longer in setup overheads than in rendering when running it through Hqueue?

[edward]

Have you tried running hython/hbatch manually on the render node to see if it's just regular process start up time vs hqueue scheduling?

[danw]

I've used hbatch via batch files before, and never had a problem - it's always been even faster and more efficient than running anything via the GUI.

 

I just tried running hython, and the running a basic hou.hipFile.load("<somefile>") command.  hython itself seemed to wait for a beat before it loaded... maybe up to a second.  The hip file loading seemed almost instant.  Not sure what to test beyond that, I'm still pretty much a novice when it comes to Python.

[Mandrake0]

it looks like there is something wrong with the network setup. maybe the routing is wrong so it takes a while till it finds the other host.