I have been testing events for ILC lately, in which both the events and the detectors are far more complicated than for HPS, and the memory usage was varying between 600 and 700 MB. I can confirm this for HPS but I am pretty sure the memory usage is reasonable (e.g. well below 1 GB per process).

LCSim probably does have some memory use inefficiencies and possibly even memory leaks that should be investigated. The basic jconsole can be used to look at memory usage on the fly. If it is discovered that lcsim/hps-java is using more memory than is considered reasonable, we should open a bug report in HPS Java and investigate it further.

On Mar 5, 2014, at 2:12 PM, Maurik Holtrop <[log in to unmask]> wrote:

Dear All,

While here at jlab, I was told in passing that our jobs on the Jlab farm are taking up too much memory. The farm systems have 2 GB per core (I think?) and some of our jobs are asking for 4 GB. That means that one job is actually using up the resources for 2.

I am wondering if this is really a required foot print for our code, or whether there are ways to reduce the memory usage. For full out data processing of our real data, this could become a critical issue.

What is the memory footprint of our reconstruction code, and what does it depend on?
What causes it to be so large, can we reduce it?

For the current MDC, perhaps it would be possible to split the SLIC step, which takes little memory, from the reconstruction step which takes a lot. This would increase the throughput on the farm, since more nodes would be available to us.

Best,
Maurik

<PastedGraphic-4.png>

Use REPLY-ALL to reply to list
To unsubscribe from the HPS-SOFTWARE list, click the following link:
https://listserv.slac.stanford.edu/cgi-bin/wa?SUBED1=HPS-SOFTWARE&A=1