uima-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Eddie Epstein <eaepst...@gmail.com>
Subject Re: DUCC- Agent1 is on Physical and Agent2 is on virtual=Slow the job process timing
Date Fri, 19 Dec 2014 14:36:24 GMT
Hi Reshu,

On Fri, Dec 19, 2014 at 12:26 AM, reshu.agarwal <reshu.agarwal@orkash.com>

> Hi,
> Is there any problem if one Agent node is on Physical(Master) and one
> agent node is on virtual?
> I am running a job which is having avg processing timing of 20 min when I
> have configured a single machine DUCC (physical machine)as well as when
> both nodes were on physical machine only.

So the job is running at the same speed on one physical machine as on two
physical machines?The two machine have similar CPU performance and number
of cores?

> When I have shifted my one agent node to virtual machine avg processing
> timing of Job was increased to 1 Hour. Here I noticed that my job driver
> was also running only on virtual machine's agent node.

Hard to diagnose with such little information. Look on the job details
page. The work items tab will show the processing time for each work item
and the machine it ran on. See if the timing is clearly different for work
on virtual machine. Look at the performance tab and compare the breakdown
between fast and slow jobs, where are the differences?

There are several factors which influence performance. Is the job process
CPU bound or does it have significant I/O wait time? Are the total number
of processing threads on a machine much more than the number of real CPU
cores on the machine? Average CPU usage for each JP are shown on the
processes tab of each job; 100% = one CPU, 200% = 2 CPU. How do these
numbers look vs the number of processing threads each JP is running?

> Can we run job driver to specific agent node so that I will be able to
> test any other Case Scenario? Because I also tried to run my job's process
> on agent node of physical machine but it didn't reflect the processing time
> much.

Normally a job driver is not a bottleneck. It can be a bottleneck if the
driver is sending raw data instead of references to raw data to the JPs. Or
if the JD is running on a machine that is in bad shape, paging, etc. What
is the CPU reported for the JD?

> Thanks in advanced.
> Reshu.

  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message