RDMA optimization on MXNet/ps-lite

• This optimization is based on MXNet 0.10
• This optimization is done on Linux Centos Sugon cluster(10 nodes).
• This optimization requires infiniband card & lmlx4 library

Attention

• ps-lite-rdma-final/ source code is different from report/final-submit/ source code. But both of them can run smmothly.
• The differences between these two source code files are:
  • ps-lite-rdma-final is completely written by Lin Zhiqi (the owner of this repository), the final-submit source code is written by Song Xiaoniu.
  • The major difference between this two source code is the basic model of RDMA QP and CQ
    • ps-lite-rdma-final uses 1 shared send cq (not srq!) on all QPs. Each QP has its own recv cq.
    • final-submit use the RDMA model that each QP has its own send cq and recv cq
    • Features ps-lite-rdma-final has but final-sbumit doesn't have:
      • Parallel memcpy (by unlocking early locks of rdma send operation)
      • multi-post-recv-request(repeatly post multi recv request at end of connection setup, thus can provide higher performance when facing with n workers - 1 server)
• These two codes have similar performance. But due to final-submit has more sample tests results, so we finally use this version to submmit the final-report.