Problem with Spectre Monte Carlo Distributed Processing

C

Chris

Guest
Hi All,
When I try to submit, say 20 Monte Carlo runs with
hostMode('distributed) which are separated into 5 parallel jobs,
whether it be with the Analog Stats GUI or within an Ocean script, the
Monte Carlo jobs run serially rather than in parallel and this sort of
defeats the whole purpose of running several jobs. What I notice by
looking at the spectre.out files is that the later jobs (like
iterations 6->10) are "waiting" for the previous job (like iterations
1->5) to finish. It seems to me that these should all be independent
of each other. If it were supposed to work like this, there would be
no point in having the distributed option and I would just run all 20
jobs in series. Can anyone tell me what I might be doing wrong. I have
been troubleshooting this for a long, long time.
Thanks in Advance!
Chris Macchietto
LSI
 
Hi,

I think one MC analysis is one job, it does not matter how many
samples you run. It may support multipule threads over multipule CPUs.

JD

On Aug 21, 3:56 pm, Chris <chris.macchie...@lsi.com> wrote:
Hi All,
When I try to submit, say 20 Monte Carlo runs with
hostMode('distributed) which are separated into 5 parallel jobs,
whether it be with the Analog Stats GUI or within an Ocean script, the
Monte Carlo jobs run serially rather than in parallel and this sort of
defeats the whole purpose of running several jobs. What I notice by
looking at the spectre.out files is that the later jobs (like
iterations 6->10) are "waiting" for the previous job (like iterations
1->5) to finish. It seems to me that these should all be independent
of each other. If it were supposed to work like this, there would be
no point in having the distributed option and I would just run all 20
jobs in series. Can anyone tell me what I might be doing wrong. I have
been troubleshooting this for a long, long time.
Thanks in Advance!
Chris Macchietto
LSI
Click to expand...
 
My observations indicate, that the first job needs to run one complete
analysis without alterations while all other simulation wait.
After this first run is completed, the altered monte carlo simulations
are calculated and all jobs start to work in parallel.

Jan



On 21 Aug., 23:56, Chris <chris.macchie...@lsi.com> wrote:
Hi All,
When I try to submit, say 20 Monte Carlo runs with
hostMode('distributed) which are separated into 5 parallel jobs,
whether it be with the Analog Stats GUI or within an Ocean script, the
Monte Carlo jobs run serially rather than in parallel and this sort of
defeats the whole purpose of running several jobs. What I notice by
looking at the spectre.out files is that the later jobs (like
iterations 6->10) are "waiting" for the previous job (like iterations
1->5) to finish. It seems to me that these should all be independent
of each other. If it were supposed to work like this, there would be
no point in having the distributed option and I would just run all 20
jobs in series. Can anyone tell me what I might be doing wrong. I have
been troubleshooting this for a long, long time.
Thanks in Advance!
Chris Macchietto
LSI
Click to expand...
 
On 24 Aug., 18:19, Geier <sunde...@gmx.de> wrote:
My observations indicate, that the first job needs to run one complete
analysis without alterations while all other simulation wait.
After this first run is completed, the altered monte carlo simulations
are calculated and all jobs start to work in parallel.
Click to expand...
With some simulators this first run can be turned off by setting
the .NONOMIN option before starting the monte-carlo. I don't know from
the top of my head if this option also exist in spectre.

--
Svenn
 
I have checked the manual.
The parameter is called "donominal". Possible values are yes and no.
It seems, that the nominal run simply is intended to check whether a
simulation completes without errors.

Thus donominal=no (don't know yet where to set in the ADE) avoids that
run.

Jan

On 25 Aug., 13:19, Svenn Are Bjerkem <svenn.bjer...@googlemail.com>
wrote:
On 24 Aug., 18:19, Geier <sunde...@gmx.de> wrote:

My observations indicate, that the first job needs to run one complete
analysis without alterations while all other simulation wait.
After this first run is completed, the altered monte carlo simulations
are calculated and all jobs start to work in parallel.

With some simulators this first run can be turned off by setting
the .NONOMIN option before starting the monte-carlo. I don't know from
the top of my head if this option also exist in spectre.

--
Svenn
Click to expand...
 
On Wed, 22 Aug 2007 16:46:06 -0000, JD <Jiandong.Ge@gmail.com> wrote:

Hi,

I think one MC analysis is one job, it does not matter how many
samples you run. It may support multipule threads over multipule CPUs.

JD

On Aug 21, 3:56 pm, Chris <chris.macchie...@lsi.com> wrote:
Hi All,
When I try to submit, say 20 Monte Carlo runs with
hostMode('distributed) which are separated into 5 parallel jobs,
whether it be with the Analog Stats GUI or within an Ocean script, the
Monte Carlo jobs run serially rather than in parallel and this sort of
defeats the whole purpose of running several jobs. What I notice by
looking at the spectre.out files is that the later jobs (like
iterations 6->10) are "waiting" for the previous job (like iterations
1->5) to finish. It seems to me that these should all be independent
of each other. If it were supposed to work like this, there would be
no point in having the distributed option and I would just run all 20
jobs in series. Can anyone tell me what I might be doing wrong. I have
been troubleshooting this for a long, long time.
Thanks in Advance!
Chris Macchietto
LSI
Click to expand...
No, this is not true if you're using distributed. You should be able to chunk
the jobs as Chris described, and they should be able to run in parallel.

I don't know why that isn't working. It's definitely worked for me in the past.

Regards,

Andrew.
--
Andrew Beckett
Senior Solution Architect
Cadence Design Systems, UK.
 
You must log in or register to reply here.

Welcome to EDABoard.com

Sponsor

Back
Top