View Single Post
Old 04-16-2014, 07:14 PM   #25
Brian Bushnell
Super Moderator
Location: Walnut Creek, CA

Join Date: Jan 2014
Posts: 2,707

Originally Posted by GenoMax View Post
Brian: I am not nitpicking but many of us run BBMap on clusters via queuing systems so setting that flag (and the threads=n) is critical (to prevent BBMap from overstepping the queue resource limits).

For clusters, I recommend setting the -Xmx flag to indicate the amount of memory you requested and 'threads' (or 't') to the number of slots you have requested. We (JGI) run on clusters too, but most of the nodes are exclusive-only, so all hardware resources are available to the user.

The new memory-detection system should work on Linux/UGE clusters (it works here). The shellscript calculates the system's free virtual memory, free physical memory, and 'ulimit -v'. Then it will invoke Java with the minimum of those 3 values, and thus, hopefully, will never fail initialization, and never use more RAM than is allowed. Although here, at least, it's impossible to use more RAM than allowed because our cluster will detect that and kill the process.

Our clusters use UGE (formerly known as SGE) and set 'ulimit' according to the qsub memory request. I don't know if that is universal or just a local policy, but at least here, users should be able to use the shellscripts without specifying -Xmx, and they should work. I'd be happy if someone elsewhere tried it without setting -Xmx and reported what happens.

Threads is a little more difficult. For now, unless you have exclusive access to a node, I recommend that you specify the number of threads allowed (e.g. t=4 if you reserved 4 slots). There is an environment variable "NSLOTS" that apparently gets set by UGE, which I will parse in my next release, and default to using the minimum of the system's reported logical cores and NSLOTS. But it appears to not always be correct, and furthermore, hyperthreading makes the matter more confusing (BBMap does not benefit much from hyperthreading, but many of the other BBTools do, like BBNorm).

Also - there are basically 4 types of programs in BBTools, resource-wise:

1) Super-heavyweight:
Requests all available threads and memory by default, but can be capped with the '-Xmx' and 't' flags. That's because all are both multithreaded and require an amount of memory dependent on the input, which the shellscript can't calculate. This includes,,,,,,,,,, and

2) Heavyweight:
Requests all available RAM but only one primary thread. Pipeline-multithreading (input thread, processing thread, output thread) cannot be disabled, so the 't' flag has no effect unless you are doing multithreaded compression using pigz. Includes,,, and

3) Midweight:
Requests all available threads by default, but can be capped with the 't' flag; needs little memory, so the -Xmx flag is hard-coded at an appropriate value that should work for all inputs (ranging from 100m to 400m) and not affected by available memory, though you can still override it. This includes

4) Lightweight:
Needs minimal memory and only one primary thread, so again the -Xmx flag is hard-coded to a low value (at most 200m) rather than dependent on autodetection. Again, 't' flag has no effect unless you are doing multithreaded compression using pigz. This includes,,,,,,,, and

Note that if you have pigz installed, you can accelerate all BBTools performance on gzipped input or output using the "unpigz" and "pigz" flags (which can be set to 't' or 'f'). pigz allows multithreaded .gz compression and decompression, and is both faster and more efficient (even with only a single thread) than Java or gzip. So with pigz installed, even a mostly singlethreaded program like can (and by default, will) use all all available threads if you write gzipped output: in=read1.fq.gz in2=read2.fq.gz out=interleaved.fasta.gz zl=8

..will compress the output to gzip compression level 8. If pigz is installed, it will use pigz for both compression and decompression instead of Java, resulting in decompression using around 1.5 cores per input file and compression using all allowed cores. You can prevent this with the 't=1' flag (to cap compression at 1 thread) or 'unpigz=f' and 'pigz=f' (to disable decompression/compression in pigz processes).

Due to some bugs in UGE (which can kill processes that fork in certain circumstances), pigz is by default set to false in high-memory processes and true in low-memory processes.

If you have not tried pigz, I highly recommend it! It's great and will become increasingly important as core counts increase. It's a drop-in replacement for gzip - same command line, 100% inter-compatible, but multithreaded and more efficient per thread. The only downside is increased memory usage, but it's not bad - around 12 MB per thread. Compression (for genomic data) is within ~0.1% of gzip at the same compression level. As I mentioned, decompression uses at most 1.5 cores in my tests, while compression can use all cores.

Last edited by Brian Bushnell; 04-17-2014 at 09:46 AM.
Brian Bushnell is offline   Reply With Quote