Skip to content

Latest commit

 

History

History
74 lines (51 loc) · 3.57 KB

README-ori.md

File metadata and controls

74 lines (51 loc) · 3.57 KB
Raw

bwa-mem2

IMPORTANT:
1.'bwa-mem2 index' requires ~48GB of space for full human genome plus index.
2. We recommend to run the code in AVX512 mode to get maximum performance.
3. The command-line of bwa-mem2 is exactly same as bwa-mem.

Compile:
$ make
'make' compiles using SSE4.1 vector mode by default.

Compile options:
make CXX=<compiler>, e.g compiler can 'g++' or 'icpc'
make arch=<mode>

mode:
a. native - generates 'bwa-mem2' binary for the architechture on which it is compiled. It detects the underlying harware flags for AVX512/AVX2/SSE2 vector modes and compiles accordingly. For example, Intel Xeon Haswell supports AVX2 vector mode; Intel Xeon Skylake supports AVX512 vector mode. AVX vector support by the processor can be checked in /proc/cpuinfo file.
b. sse - generates 'bwa-mem2' SSE4.1 vector code binary
c. avx2 - generates 'bwa-mem2' AVX2 vector code binary
d. avx512bw - generates 'bwa-mem2' AVX512BW vector code binary

make multi
Generates binaries for all the three vector modes: bwa-mem2.sse41, bwa-mem2.avx2, bwa-mem2.avx512bw and also generates a binary called bwa-mem2 that, if run, detects the underlying architecture and runs the corresponding binary out of bwa-mem2.sse41, bwa-mem2.avx2, bwa-mem2.avx512bw.

Run:

  • Create the index: 'bwa-mem2 index '
  • Mapping: 'bwa-mem2 mem' can be run with appropriate paramters
    e.g. bwa-mem2 mem -t 1 <read1.fq> [<read2.fq>] > <out.sam>

Notes:

  • In SSE4.1 vector mode, bwa-mem2 uses bwt index with 2-bit representation
  • In AVX512/AVX2 vector mode, bwa-mem2 uses bwt index with 8-bit representation

Example run:
sh bwa-mem2 index datasets/hg38.fa
bwa-mem2 mem -t 1 datasets/hg38.fa datasets/SRR099967.filt.fastq > datasets/aln-se.sa

NUMA:
In multi-socket system, it is beneficial to bind the process memory to a socket
Linux command numactl can be used to bind process memory to a given numa domain
For exmaple, the following command binds the memory to socket-0
$ numactl -m 0 bwa-mem2 mem -t 10 -o datasets/aln-se.sa datasets/hg38.fa datasets/SRR099967.filt.fastq

Similarly, the process threads can be bound to cores.
$ numactl -m 0 -C 0-9 bwa-mem2 mem -t 10 -o datasets/aln-se.sa datasets/hg38.fa datasets/SRR099967.filt.fastq

$ lscpu - Linux command provides the architectural details such as list of NUMA domains and their CPU lists

Directory structure (bwa-mem2):

  • src/ - contains source code files
  • test/ - contains test files to run individual benchmarks for smem, sal, bsw (disabled for now)
  • ext/ - contains external Intel safestring library for secure string operations
  • images/ - contains bwa-mem2 perfomance charts, demonstrating its speedup against original bwa-mem

Notes:

  • As mentioned previously, the code supports four exceution modes:

    1. Vector SSE2: It runs bwa-mem2 with the kernels running SSE2 vector instructions (128 bits vector-width)
    2. Vector AVX2: It runs bwa-mem2 with the kernels running AVX2 vector instructions (256 bits vector-width)
    3. Vector AVX512: It runs bwa-mem2 with the kernels running AVX512 vector instructions (512 bits vector-width)

  • BWA-MEM2 displays run-time performance profiling of the code on standard output. (Hopefully, we managed it to be self-explanatory)