The language here indicates some confusion. There is no "alternate GSL". We use the ROOT MathMore interface. Some concrete algorithms implementing that interface come from GSL, but not all. In particular the adaptive algorithm, I think, does not come from GSL. But GSL algorithms are typically the best ones. Speed is one aspect only, and we can not tune it independently of accuracy. From memory, what Igor (or was it Cezary?) was talking about was a variable transformation that removes a discontinuity in a simple integration space.
cheers
Costas
Sent from my iPhone
On 3 Aug 2016, at 21:05, Steven Dytman <[log in to unmask]<mailto:[log in to unmask]>> wrote:
maybe in his note from last week?
Otherwise, good to ask him as he's often responsive.
Steve
On 8/3/2016 2:01 PM, Gabriel Nathan Perdue wrote:
Did he mention anything about how to enable that? We can figure it out, but it would be nice to have a recipe, etc.
pax
Gabe
On Aug 3, 2016, at 12:59 PM, Steven Dytman <[log in to unmask]<mailto:[log in to unmask]>> wrote:
don't forget Igor's suggestion that alternate GSL inside
root is a source of large increase in speed.
On 8/3/2016 1:44 PM, Gabriel Nathan Perdue wrote:
This fits with anecdotal observations of my own. In particular, the charm DIS stuff from out standard validation has never finished (up to 24 hour runs), and the DIS stuff always finished last.
We did switch away from LHAPDF to use a custom PDF - perhaps this is the issue? Some subtlety in the implementation?
pax
Gabe
On Aug 3, 2016, at 12:34 PM, Robert W Hatcher <[log in to unmask]<mailto:[log in to unmask]>> wrote:
Hi all,
Per the ongoing discussion of slowness in GENIE spline generation I ran some test samples with trunk (as of 2016-07-18) and R-2_10_6 -- both using the same underlying support code, e.g. ROOT, GSL....
I used my DAG generating script which runs in stages:
1) generate { nu nu-bar } x { e mu tau } x { free-n free-p } = 12 jobs
2) combine those
3) generated all 6 nu flavors on a single (A,Z) isotope = 106 jobs
For these tests I generated only sub-sets of the EventGeneratorList: RES DIS QE CCMEC
and for 100 knots up to 400 GeV.
Here's a link to my evernote with numbers:
https://www.evernote.com/l/AQoHQHYmOyVIGYrAi_tnzj4-MwsW9sMsCPw<https://na01.safelinks.protection.outlook.com/?url=https%3a%2f%2fwww.evernote.com%2fl%2fAQoHQHYmOyVIGYrAi_tnzj4-MwsW9sMsCPw&data=01%7c01%7cdytman%40PITT.EDU%7c50853381bd824b9796ee08d3bbc5c95a%7c9ef9f489e0a04eeb87cc3a526112fd0d%7c1&sdata=JI3Tlzcd7gP2CemIB2F3ChmpPkFcTXkDLfTYAWMn3nI%3d>
major standouts (# = sum of wall clock time in seconds):
RES stage 1 (trunk) 237134 / (r2106) 341309 = 0.7 e.g. faster
DIS stage 1 (trunk) 687672 / (r2106) 209392 = 3.28 e.g. slower
QE stage 3 (trunk) 51753 / (r2106) 107353 = 0.48 e.g. faster
CCMEC stage 3 (trunk) 101307 / (r2106) 86677 = 1.17 e.g. slower
This seems to implicate DIS generation as the big problem. I don't remember any big changes to that code per se. Does trunk switch away from using LHAPDF for the parton distributions? Perhaps that's a performance issue?
-robert
Robert Hatcher
Computational Physics Developer
Scientific Computing Division
Fermi National Accelerator Laboratory
PO Box 500, MS 234, Batavia, IL 60510, USA
[log in to unmask]<mailto:[log in to unmask]>
Office: 630-840-3102<tel:630-840-3102>
Cell: 630-234-0091<tel:630-234-0091>
Gabriel Perdue
Associate Scientist
Scientific Computing Division
Fermi National Accelerator Laboratory
PO Box 500, MS 234, Batavia, IL 60510, USA
[log in to unmask]<mailto:[log in to unmask]>
Office: 630-840-6499
Cell: 630-605-8062
Connect with us!
Newsletter<https://na01.safelinks.protection.outlook.com/?url=http%3a%2f%2fnews.fnal.gov%2ffermilab-at-work%2f&data=01%7c01%7cdytman%40PITT.EDU%7c50853381bd824b9796ee08d3bbc5c95a%7c9ef9f489e0a04eeb87cc3a526112fd0d%7c1&sdata=wLmn7O096uyHPqtQuAFXPug4EYo%2f2Wktj7xXU0iF56w%3d> | Facebook<https://na01.safelinks.protection.outlook.com/?url=http%3a%2f%2fwww.facebook.com%2fFermilab&data=01%7c01%7cdytman%40PITT.EDU%7c50853381bd824b9796ee08d3bbc5c95a%7c9ef9f489e0a04eeb87cc3a526112fd0d%7c1&sdata=vQ0xmVW3%2fVWy4tVNYuy%2fnWxqe2CasqEJQbFKoXz15%2bQ%3d> | Twitter<https://na01.safelinks.protection.outlook.com/?url=https%3a%2f%2fwww.twitter.com%2ffermilab&data=01%7c01%7cdytman%40PITT.EDU%7c50853381bd824b9796ee08d3bbc5c95a%7c9ef9f489e0a04eeb87cc3a526112fd0d%7c1&sdata=jmvtvHcnQkobHhtuGy4pzDHpEKHcG4WQMVzJsJBg1Tc%3d>
Gabriel Perdue
Associate Scientist
Scientific Computing Division
Fermi National Accelerator Laboratory
PO Box 500, MS 234, Batavia, IL 60510, USA
[log in to unmask]<mailto:[log in to unmask]>
Office: 630-840-6499
Cell: 630-605-8062
Connect with us!
Newsletter<https://na01.safelinks.protection.outlook.com/?url=http%3a%2f%2fnews.fnal.gov%2ffermilab-at-work%2f&data=01%7c01%7cdytman%40pitt.edu%7c2965070c6e1a4b12353e08d3bbc83ab9%7c9ef9f489e0a04eeb87cc3a526112fd0d%7c1&sdata=IHHVQQe7rVTCppRsG0VLm05jyOZtypAJRYILdSwoKh8%3d> | Facebook<https://na01.safelinks.protection.outlook.com/?url=http%3a%2f%2fwww.facebook.com%2fFermilab&data=01%7c01%7cdytman%40pitt.edu%7c2965070c6e1a4b12353e08d3bbc83ab9%7c9ef9f489e0a04eeb87cc3a526112fd0d%7c1&sdata=WWYby3OjUlz%2bAfqOpQnTgHc1vYvrvWJptX2B5PQ5Ly0%3d> | Twitter<https://na01.safelinks.protection.outlook.com/?url=https%3a%2f%2fwww.twitter.com%2ffermilab&data=01%7c01%7cdytman%40pitt.edu%7c2965070c6e1a4b12353e08d3bbc83ab9%7c9ef9f489e0a04eeb87cc3a526112fd0d%7c1&sdata=7F8J90Vy6UgeCnxpN%2bO4XRjTGgK55ioJm5y%2fkUfk7X0%3d>
|