Release Notes

[Siddharth Srinivasan]

It would be very useful to have detailed release notes for each new release, so people are aware of the specific differences between the various versions, and whether or not to upgrade. For example the latest 3.1 release says "This release includes major performance improvements, especially to the AMOEBA force field, as well as support for parallelizing computations across multiple GPUs, bug fixes, and more", but as a user I have no idea if

1. Parallelizing across multiple GPUs: does this work on OpenCL and CUDA both? What is the scaling performance like on your test cases, i.e do I get 2X the speedup using 2 GPUs or 1.5X?
2. What bugs were fixed, do they affect any of my current work?
3. What other changes/improvements have been made, which are relevant to my decision to upgrade versions?

This would be a very useful document to maintain. Thanks for the great work, all!

[Peter Eastman]

I see you've noticed the new release! There should be an official announcement going out soon which will have more details about the new features. In the mean time, I'll try to answer your questions.

> 1. Parallelizing across multiple GPUs: does this work on OpenCL and CUDA both? What is the scaling performance like on your test cases, i.e do I get 2X the speedup using 2 GPUs or 1.5X?

It's OpenCL only. Using it is very simple: use the OpenCLDeviceIndex property as before, but now it allows you to specify a comma separated list of device indices:

properties["OpenCLDeviceIndex"] = "0,1"

The performance won't be linear in the number of GPUs, so a 1.5x speedup with two GPUs is probably close to what you'll get.

> 2. What bugs were fixed, do they affect any of my current work?

If you've entered a bug report into the tracker and it was fixed in this release, the report should have been closed.

> 3. What other changes/improvements have been made, which are relevant to my decision to upgrade versions?

The list of major changes is on the roadmap:

http://wiki.simtk.org/openmm/RoadmapTimeline

I definitely recommend upgrading. The performance improvements and bug fixes alone are sufficient reason.

Peter

[John Chodera]

I'll second Siddharth's request to have a list of bugs fixed in each release. One of the major issues we had with gromacs was that it was difficult to figure out which bugs were fixed in which releases, which became especially problematic when we had to determine whether the new release meant we had to repeat calculations whose results may have changed or (in the worst case) retract publications. For a while, it was possible to determine this from their changelog (manually correlating this with release histories), but then they stopped maintaining this. Having a nice list of which bugs were fixed and which, if any, calculations might have been affected, would be very nice for your users.

[Siddharth Srinivasan]

Thanks for the information Peter.

I should note, I also posted this at the new beta forum I was instructed to use, at https://simtk.org/forums/viewtopic.php?f=161&t=3055 but looks like this old forum is being monitored more actively? I also posted another question at the new forum, so I am linking it here. See https://simtk.org/forums/viewtopic.php?f=161&t=3065

[Joy Ku]

Hi Siddharth,

Thanks for testing out the test version of our new forums. I hope you've found it to be a vast improvement over our current version.

I apologize if our message about the test forums wasn't clear, but that is the test version, intended to work out any last bugs and as you pointed out, not actually being monitored at this point. Please continue posting to this version of the forum. I would also suggest that you repost your new question to this forum as well, as the test forum is a temporary sandbox from which data will be deleted.

If you are able to continue exploring the test forums, we would appreciate any additional feedback you might have.

Thank you.

Best,
Joy

[Taras Y]

Hi Peter, I would like to shortly come back to the question of supporting multiple GPUs.

The ability of asking for more than 1 GPU in OpenCL via properties["OpenCLDeviceIndex"] = "0,1, ...":
how does this correlate with your earlier comment that OpenMM does not make any attempt to use multiple GPUs in one simulation?

Also: do you have plans to extend OpenMM to support multiple GPUs with technologies like gpudirect 2.0? How complicated might this be (e.g. how many programmer's hours this might take)?

Thank you! Regards, Taras.

[Peter Eastman]

Hi Taras,

Using multiple GPUs for one simulation is a new feature in 3.1. When I previously said OpenMM didn't use multiple GPUs, I was talking about an earlier version that did not have this ability.

Currently, the GPUDirect features are only accessible from CUDA, not OpenCL. If Nvidia adds an extension for using them from OpenCL, we'll definitely look into using them.

Peter