InfiniBand and 10-Gigabit Ethernet for Dummies
InfiniBand and 10-Gigabit Ethernet for Dummies
InfiniBand and 10-Gigabit Ethernet for Dummies
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Time (us)<br />
3000<br />
2500<br />
2000<br />
1500<br />
<strong>10</strong>00<br />
500<br />
MPI-Level Two-sided Communication<br />
0<br />
Memcpy+Send<br />
MemcpyAsync+Isend<br />
MVAPICH2-GPU<br />
32K 64K 128K 256K 512K 1M 2M 4M<br />
Message Size (bytes)<br />
with GPU Direct<br />
• 45% <strong>and</strong> 38% improvements compared to Memcpy+Send, with <strong>and</strong> without<br />
GPUDirect respectively, <strong>for</strong> 4MB messages<br />
• 24% <strong>and</strong> 33% improvement compared with MemcpyAsync+Isend, with <strong>and</strong> without<br />
GPUDirect respectively, <strong>for</strong> 4MB messages<br />
Time (us)<br />
3000<br />
2500<br />
2000<br />
1500<br />
<strong>10</strong>00<br />
500<br />
0<br />
Memcpy-Send<br />
MemcpyAsync+Isend<br />
MVAPICH2-GPU<br />
32K 64K 128K 256K 512K 1M 2M 4M<br />
Message Size (bytes)<br />
without GPU Direct<br />
H. Wang, S. Potluri, M. Luo, A. Singh, S. Sur <strong>and</strong> D. K. P<strong>and</strong>a, MVAPICH2-GPU: Optimized GPU to<br />
GPU Communication <strong>for</strong> <strong>InfiniB<strong>and</strong></strong> Clusters, ISC ‘11<br />
NVIDIA-SC '11<br />
12