SoftwareMogulRelease2014NFSissuesTest

Issues with Mogul from CSDS 2014

Mogul speed tests

• As described on previous page SoftwareMogulRelease2014NFSissues there as issues using mogul from grade, grade_PDB_ligand and buster-report.
• Here we describe tests comparing the cpu timing from mogul as invoked by grade_PDB_ligand.
• Stop press! There is now an update that has much better speed, to download http://www.ccdc.cam.ac.uk/products/csd_system/updates
• To provide a range of typical ligands for macromolecular crystallography 496 PDB ligands with three letter codes in the range 000 to 587.
• The tests compare using 2 versions of mogul either installed on a fast (SSD) local disk or from an NFSv4 mounted file system.
• grade_PDB_ligand was run four times for each ligand with the different mogul executable/mount method. The CPU timings are as reported in perl as "real" CPU seconds (this should correspond to elapsed).
• The tests are run on a Intel(R) Xeon X5660 with 12 cores 24 threads running at 2.80GHz. NFS is over gigabit ethernet,

2014 CSDS mogul is on average 2.5 times as slow as 2013

Comparing 2014 Mogul installed on local disk with 2013 Mogul installed on local disk. Conclusion 2014 Mogul is on average 2.5 times as slow as 2013. But cpu still in seconds and quite usable by grade and buster report

For most ligands there is no problem using mogul from NFS (v4)

• Looking at difference in cpu speeds for Mogul installed on NFS (v4) compared to a local disk:

For the great majority of ligands it makes little difference whether Mogul is installed on NFS (v4) or a local disk.

But there are outliers where using CSDS NFS (v4) results in slow runs

The same graph as previous but expanded to show the outliers

• Note that for CSDS2013 there where outliers where NFS resulted in a much slower run (up to 120 secs for NFS compared to maximum of 20 seconds for a local installation).
• The slower mogul from CSDS 2014 means that outliers use up to 430 seconds cpu compared to maximum 50 seconds cpu for a local installation.

Which ligands show the NFS slow down problem?

• Here is a selection of the ligands which were slow:

ligand id	ligand image	CPU seconds 2013local,2014local,2013nfs,2014nfs
548	(Pfizer compound)	11.9,37.2,58.2,425.1
287	(Bristol-Myers Squibb)	12.3,39.6,55.5,408.4
219		7.8,28.4,120.4,330.6
205		9.9,23.3,105.6,317.5
STU		11.1,37.7,150.0,593.7

• last ligand STU was the one that originally came to our attention rather than from the test - mogul run using CSDS 2014 is really slow (10 minutes cpu).
• Do not know what the common factor - some structures have many rings, some have "exotic" atoms???
• If you want to get your own mogul timings for PDB ligands then this can be done simply see SoftwareMogulRelease2014NFSissueOwnTest

Page original version 7th January 2014, Updated 1st May 2014. Address problems, corrections and clarifications to buster-develop@globalphasing.com