<Design and evaluation of benchmarks for financial applications using Advanced Message Queuing protocol (amqp) over InfiniBand>
Hari subramoni, Gregory Marsh, sundeep narravula, Ping Lai, and dhabaleswar K. Panda
Department of Computer Science and Engineering, the Ohio State University
InfiniBand is a new network transmission technology in recent years. It features high bandwidth and low latency. A unified interconnection structure is formed through a persistent cable connection mode, which can process storage I/O, network I/O, and inter-process communication (IPC ). It can eliminate the bottlenecks that currently impede servers and storage systems. It is a high-performance I/O Technology Dedicated to servers rather than PCs and is mainly used in data centers.
InfiniBand supports two transmission models: Channel semantics and memory semantics ). The main channel mode is discrete data transmission and receiving, and the memory mode uses rdma technology to allow the process to access the memory of the remote process without passing through the CPU. InfiniBand supports the TCP/IP protocol, IP over Ib (ipoib), and socket direct protocol (SDP ).
Amqp performance test
Test environment:
Node:
CPU: Intel Xeon quad dual-core processor, 6 GB RAM
Operating System: Red Hat Enterprise Linux 4u4.
NIC: 1 gige network interface controller (NIC) with InfiniBand Host channel adapter (HCA)
Broker version
Qpid version m3 alpha
Running parameter: TCP no delay
Test the basic environment:
Test the maximum network traffic (using socket ):
Ipoib: 550 Mbps
SDP: 650 Mbps
Direct Exchange-single publisher single consumer (DE-SPSC)
(A) DE-SPSC small message latency
DE-SPSC large message latency
DE-SPSC message rate, and
(D) MPI level message Rate
Direct Exchange-multiple publishers multiple consumers (DE-MPMC)
(A) DE-MPMC bandwidth
(B) DE-MPMC CPU utilization over ipoib
(C) DE-PP small message latency
DE-PP large message latency
Fanout exchange-single publisher multiple consumers (FE-SPSC)
(A) bandwidth
(B) ipoib message Rate
(C) 1 gige message Rate
(D) SDP message Rate
Topic exchange-single publisher single consumer (TE-SPSC) Benchmark
(A) bandwidth
(B) ipoib message Rate
(C) 1 gige message Rate
(D) SDP message Rate
Conclusion:
Ipoib has better processing performance for small messages and SDP has better performance for large data processing.
The CPU usage of borker is proportional to the number of producer and consumer.
Ipoib is very sensitive to message rate, which is related to the stack overhead of ipoib. If rdma technology is used, the performance will be further improved.
Test data shows that it is best to mix these technologies and add multiple brokers to solve the broker bottleneck problem.