<html>
<head>
<style><!--
.hmmessage P
{
margin:0px;
padding:0px
}
body.hmmessage
{
font-size: 10pt;
font-family:Tahoma
}
--></style></head>
<body class='hmmessage'><div dir='ltr'>
Hi Gus,<div>The reason I'm interested to have a local copy in each node is because I am working to compare analysis performance between message passing and shared memory programming model. I am trying to avoid NFS delays against the message passing model.<br>One question I have regarding your email is when you said: "hassle of copying and maintaining exact replicas on the compute nodes"</div><div>I just setting up an lab environment, not to be connected to a real network, I am not planning to do any upgrades, so what will be the reason the replicas will not be the same.</div><div>Thanks,</div><div>Sandra Guija<br><br><div><div id="SkyDrivePlaceholder"></div>> Date: Fri, 5 Oct 2012 18:13:12 -0400<br>> From: gus@ldeo.columbia.edu<br>> To: mpich-discuss@mcs.anl.gov<br>> Subject: Re: [mpich-discuss] Mpich2 - in Red Hat 7.3<br>> <br>> On 10/05/2012 03:35 PM, Pavan Balaji wrote:<br>> ><br>> > On 10/05/2012 02:33 PM, Sandra Guija wrote:<br>> >> If I decide to setup my cluster so that each node has their own mpich2<br>> >> binaries,<br>> >> it will be ok if I installed just in one and the copied the<br>> >> mpich2-install directory to the other node?<br>> >> Or I need to install idnivdually<br>> ><br>> > It's theoretically OK to just copy the install directory, but you have<br>> > to be very careful to maintain correct permissions, and directory paths,<br>> > etc. This is often a cause of many errors.<br>> ><br>> > -- Pavan<br>> ><br>> <br>> Hi Sandra<br>> <br>> As Pavan said, you can copy over the install directory to the<br>> nodes, with caution to the details that Pavan pointed out.<br>> <br>> Also, make sure all computers have the same architecture also,<br>> i.e., say all 32-bits (x86), or all 64-bits (x86_64).<br>> <br>> However, I believe it is much simpler for you to install mpich2 only<br>> in the cluster's head node, on a NFS (network file system)<br>> shared directory.<br>> I.e., install in a directory on the<br>> head node that is exported to the compute nodes<br>> and mounted on the compute nodes.<br>> Typically people create a directory only for this purpose,<br>> called, say, /shared, and you could install mpich2,<br>> say, in /shared/mpich2 or similar.<br>> <br>> The head node is the NFS server, the compute nodes are<br>> the NFS clients.<br>> <br>> This will require a single install of mpich2.<br>> You will avoid the hassle of copying and maintaining<br>> exact replicas on the compute nodes.<br>> In small clusters this is typically what people do.<br>> <br>> NFS is not much of an mpich2 list issue, but anyways ...<br>> First check if you have nfs packages (if you're on Fedora use yum)<br>> 'yum list |grep nfs'<br>> and install the packages with 'yum install' if you don't have them.<br>> Second if you already have NFS packages, check if they're<br>> running on the head node:<br>> 'service nfs status'.<br>> If NFS is not running, try<br>> 'chkconfig --list |grep nfs'<br>> If NFS is turned off, turn it on, and start the service:<br>> 'chkconfig nfs on'<br>> 'service nfs start'<br>> <br>> If you are not familiar to NFS,<br>> and how to export and mount a directory,<br>> here is the Linux NFS 'HowTo':<br>> <br>> http://nfs.sourceforge.net/nfs-howto/<br>> <br>> I hope this helps,<br>> Gus Correa<br>> _______________________________________________<br>> mpich-discuss mailing list mpich-discuss@mcs.anl.gov<br>> To manage subscription options or unsubscribe:<br>> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss<br></div></div> </div></body>
</html>