Fri Dec 03, 2021 3:43 am
Login Register Lost Password? Contact Us


Maximizing thor's performance

Comments and questions related to the Enterprise Control Language

Tue Oct 25, 2011 8:13 pm Change Time Zone

I am using a thor cluster with 5 nodes using the default configuration. I noticed that when I spray a file (big one around 4GB) it only uses 2 nodes. Also even if I distribute it, the computations are happening on two nodes only. I suspect that the configuration decides to automatically allocate only 2 slaves and keep the other 3 for other tasks.

So suppose that I am the only user of thor when I want maximum utilization of nodes and CPU, how should I configure it?
nvasil
Community Advisory Board Member
Community Advisory Board Member
 
Posts: 105
Joined: Mon Oct 17, 2011 6:48 pm

Wed Oct 26, 2011 2:31 pm Change Time Zone

There's two possibilities here:

1. Your thor has two slaves, and the other 3 nodes in your system were used for other components of the HPCC system (dali, esp, eclccserver etc). When you have a large system it's appropriate to give each of these services its own node, but on a smaller system it may be wasteful.

2. Your data is skewed so that the distribute only placed rows onto two slaves even though there are 5 slaves on the thor. I don't know how you did the distribute but if for example you did DISTRIBUTE(mydataset, myfield) where myfield had very few distinct values you can get a skew like this. However since you say that the initial spray only put data on 2 nodes it sounds more likely that it's the first case...

How did you do the configuration - using the wizard mode? And what version of the platform are you using?
richardkchapman
Community Advisory Board Member
Community Advisory Board Member
 
Posts: 110
Joined: Fri Jun 17, 2011 8:59 am

Wed Oct 26, 2011 2:37 pm Change Time Zone

Thanks for the response.
I just run the configuration tool and chose the default values. So if I have 5 nodes and I am the unique user, I will not spray files while submitting a query, in fact I will submit one query at a time, what is the best configuration so that I maximize the performance?
nvasil
Community Advisory Board Member
Community Advisory Board Member
 
Posts: 105
Joined: Mon Oct 17, 2011 6:48 pm

Wed Oct 26, 2011 2:43 pm Change Time Zone

I would probably put all my system services and the thormaster on node 1 and create 4 thor slaves on the other 4 nodes.

It _might_ be faster with thor slaves on all 5 nodes, but you'd be in danger of overloading the node that had all the system servers on.
richardkchapman
Community Advisory Board Member
Community Advisory Board Member
 
Posts: 110
Joined: Fri Jun 17, 2011 8:59 am

Wed Oct 26, 2011 2:46 pm Change Time Zone

Thanks a lot.

Now if I have a 5 node cluster with 2 slaves and I shut it down. Then I change the configuration to 4 slaves and restart it. Will everything be ok?
nvasil
Community Advisory Board Member
Community Advisory Board Member
 
Posts: 105
Joined: Mon Oct 17, 2011 6:48 pm

Wed Oct 26, 2011 3:28 pm Change Time Zone

The file meta data is stored in Dali, if you reconfigure the location of that component, then you will lose that data, or you will need manually migrate dali's meta data to the new node.

But assuming Dali stays put, and if you you reconfigure other components after ensuring all components are stopped first, you should be fine.

However, any existing data on the thor cluster(s), will still reside across the original nodes, which is also fine, but means if reading those files, you may want to DISTRIBUTE them early on in a query.
jsmith
Community Advisory Board Member
Community Advisory Board Member
 
Posts: 81
Joined: Tue Jul 19, 2011 12:58 pm


Return to ECL

Who is online

Users browsing this forum: No registered users and 1 guest

cron