Out of memory errors when running Bowtie on cluster

[biocomputer]

I'm trying to use a computer cluster for the first time, SHARCNET using the orca cluster, and I'm getting errors when trying to run Bowtie with 8 threads. 2 and 4 threads on the cluster run fine and I can run with 8 threads locally in Windows without any problems. Here are the Bowtie errors from an 8 thread run on the cluster, after which it exits:

Time loading forward index: 00:00:50
Time loading mirror index: 00:00:31
Warning: Exhausted best-first chunk memory for read HWI-ST724:196:C0LCGACXX:6:1101:1401:1999 1:N:0:ACAGTG (patid 5); skipping read
Warning: Exhausted best-first chunk memory for read HWI-ST724:196:C0LCGACXX:6:1101:1350:1984 1:N:0:ACAGTG (patid 3); skipping read
Warning: Exhausted best-first chunk memory for read HWI-ST724:196:C0LCGACXX:6:1101:1378:1995 1:N:0:ACAGTG (patid 4); skipping read
Warning: Exhausted best-first chunk memory for read HWI-ST724:196:C0LCGACXX:6:1101:1266:1963 1:N:0:ACAGTG (patid 2); skipping read
Warning: Exhausted best-first chunk memory for read HWI-ST724:196:C0LCGACXX:6:1101:1718:1950 1:N:0:ACAGTG (patid 6); skipping read
Warning: Exhausted best-first chunk memory for read HWI-ST724:196:C0LCGACXX:6:1101:1521:1954 1:N:0:ACAGTG (patid 7); skipping read
Warning: Exhausted best-first chunk memory for read HWI-ST724:196:C0LCGACXX:6:1101:1440:1961 1:N:0:ACAGTG (patid 1); skipping read
Warning: Exhausted best-first chunk memory for read HWI-ST724:196:C0LCGACXX:6:1101:1350:1929 1:N:0:ACAGTG (patid 0); skipping read
/var/spool/torque/mom_priv/jobs/3875588.orc-admin2.orca.sharcnet.SC: line 3:  3411 Segmentation fault      /work/user/bowtie/bowtie -t -p 8 -n 3 -l 75 --best --strata -m 1 --chunkmbs 8192 --suppress "1,6,7,8" RepeatMaskerClass /work/user/sequences/h3_p05_ko_combined.fa ./h3_p05_ko8c_repeatmasker.map

Here are the errors I get from the server:

--- SharcNET Job Epilogue ---
              job id: 3875588
         exit status: 139
            cpu time: 44s / 64.0h (0 %)
        elapsed time: 82s / 8.0h (0 %)
      virtual memory: 2.7G / 20.0G (13 %)

WARNING: Job died due to SIGSEGV - Invalid memory reference
WARNING: Job only used 0% of its requested walltime.
WARNING: Job only used 0% of its requested cpu time.
WARNING: Job only used 13% of its requested memory.

Locally I'm using --chunkmbs 512, Bowtie uses about 5 GB of RAM, and it starts outputting alignments. On the server I've tried up to --chunkmbs 16384 and specifying up to 24GB of RAM for the job and not even a single alignment is output before Bowtie exits with errors. I have 60million 100bp single end reads. The other Bowtie options in all cases are: -n 3 -l 75 --best --strata -m 1.

[Edit] After playing around with some more options it might be working now. I left --chunkmbs lower (1024) but requested more RAM (24GB) and it's running with no errors so far. Though if anyone can explain what's going on with memory usage here that'd be appreciated

[dpryan]

By increasing the chunkmbs option, you're telling bowtie to use more memory, which you would need to allocate on the cluster when you start the job. By not doing that, it tries to get more memory but can't. While it should catch this itself, many of us have the unfortunate habit of not checking things after malloc() and realloc(), so the program presumably doesn't know that it didn't receive the extra memory it requested. As a conseqeunce, it started trying to read/write from memory either out of range or that it otherwise didn't have access to (this is what a segfault is).

BTW, you might want to use bowtie2. While it's slightly slower it generally produces better results.

[biocomputer]

Thanks for the reply. I'm planning to switch to bowtie2 just finishing off an old project that was all done with the original bowtie.