New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Make workspaces in HGCDigitizer only depend on signal size #16336
Make workspaces in HGCDigitizer only depend on signal size #16336
Conversation
@cmsbuild please test |
The tests are being triggered in jenkins. |
A new Pull Request was created by @lgray (Lindsey Gray) for CMSSW_8_1_X. It involves the following packages: SimCalorimetry/HGCalSimProducers @cmsbuild, @civanch, @mdhildreth, @davidlange6 can you please review it and eventually sign? Thanks. cms-bot commands are listed here #13028 |
assign reco |
oh well, I guess I am not special enough |
assign reconstruction |
Comparison job queued. |
+1 |
This pull request is fully signed and it will be integrated in one of the next CMSSW_8_1_X IBs (tests are also fine). This pull request requires discussion in the ORP meeting before it's merged. @slava77, @davidlange6, @smuzaffar |
@davidlange6 please wait for Slava to sign off too. Thanks! |
On 10/25/16 9:33 AM, Lindsey Gray wrote:
I'm traveling today.
|
This one isn't incredibly urgent, so that is fine. On Tue, Oct 25, 2016 at 1:47 PM, Slava Krutelyov notifications@github.com
|
assign reconstruction |
My guess is that the size of the input MinBias matters (my test had 4 input minbias files from relval, with net total of about 30K events). |
Hi Slava, the quoted decrease is for step2! Where in step3 is using most of the 62GB? How many threads? I was using the full list of minbias files from pre15 D3Timing. |
On 10/26/16 6:39 AM, Lindsey Gray wrote:
16 threads step2 was up to 24GB, which is somewhat consistent with your "before"
|
step2 PU200 in 22624.0 memory based on MemoryCheck (post-event memory), using 16-thread job (event number reference is weak, given that it's detected by "Begin processing" and not actual event end):
So, whatever was done here to minimize memory does not show up in this metric. |
+1
looking at the same 22624 with PU200 100 events, I do not see any significant changes, beyond random variations. ... there is some decrease in the number of hits in BH, FH and EE, in the range of 1-2%, primarily at low energy, but it also matches generalTracks yield change. This is apparently random. The same signal events were used (from the logs), but minbias mix sequence was apparently different (speculation, I didn't check it, but what else). |
This pull request is fully signed and it will be integrated in one of the next CMSSW_8_1_X IBs (tests are also fine). This pull request requires discussion in the ORP meeting before it's merged. @slava77, @davidlange6, @smuzaffar |
about step3 memory, looking with single core:
So, mix replay appears to be the most costly. Writing FEVTDEBUGHLT is the next in line. The latter has cost impact both on run time memory and output on disk. |
+1 |
Change the HGCal digitizers such that only bx-vectors that contain signal are stored.
In the case of an empty detid, a single zeroed out bx-vector is used.
This reduces memory in 0 PU from avg 6.4 GB in 4-thread jobs to 2.5 GB.
This reduces memory in 200PU from peak 9.2GB in 4-thread jobs to 7.2 GB.
@slava77 would you mind running your usual tests on this PR to make sure things are OK, despite it being a SIM-only PR. I've checked to make sure everything looks sane but would like some deeper analysis.