Message: Re: superslow make on a cluster over the network Not Logged In (login)
 Next-in-Thread Next-in-Thread
 Next-in-Forum Next-in-Forum

None Re: superslow make on a cluster over the network 

Forum: Installation and Configuration
Re: Sad superslow make on a cluster over the network (Valery Taranenko)
Date: Nov 04, 03:28
From: Ben Morgan <Ben Morgan>

Hi Valery,
On Tue, 03 Nov 2009 22:36:38 GMT, Valery Taranenko wrote:
>...
> normally, on a computer with a local HDD, the make (of a 20 user files)
> w/ link takes 1 min. now, on a cluster, with a remote disk storage, i
> have to ssh to a computational node, and then the make takes 30 min or
> more. the problem is--i suspect--is a network load (I/O). "Making
> dependency for file ..." takes several min for each file(!). i suspect,
> make needs to transfer many files over the network (it's Gb Ethernet).
> not only making dependencies takes long, compilations do too. the make
> process is awfully slow. i cannot stand it :(
>...

If you're running on a computational node are there other jobs running on it? Also, it may be that there's a high network throughput anyway with lots of jobs. Both would certainly slow things down, and that being the case there isn't really a problem with Geant4 per se - it's going to happen with any buildsystem.

> does anybody knows how to bring the speed back to 1 min? i can do the
> so-called prestaging (copping the files to a local note hdd and run from
> there) but in this case all paths will be different, it will be
> basically a new g4 installation, which is too much.
<...

"Prestaging" is what I would recommend - but I'm a little unclear on what you mean by the paths being different. What you can do is in your Makefile (it sounds like you're using the Geant4 makefile system) is:

  • set G4WORKDIR to something like /tmp/geant4 - that should be on a locally mounted filesystem on *nix systems
  • set G4EXLIB to 'nothing' so that the application gets built directly without an intermediate library, i.e G4EXLIB :=

This should ensure that all created files are written locally. You'll have to copy the executable from wherever it is in /tmp to somewhere where it's available across the cluster - you can always add an 'install' target to the makefile to do this.

Cheers,

Ben.

Inline Depth:
 1 1
 All All
Outline Depth:
 1 1
 2 2
 All All
Add message: (add)

1 Idea: Re: superslow make on a cluster over the network   (Valery Taranenko - Nov 04, 17:34)
(_ None: Re: superslow make on a cluster over the network   (Ben Morgan - Nov 05, 03:01)
 Add Message Add Message
to: "Re: superslow make on a cluster over the network"

 Subscribe Subscribe

This site runs SLAC HyperNews version 1.11-slac-98, derived from the original HyperNews


[ Geant 4 Home | Geant 4 HyperNews | Search | Request New Forum | Feedback ]