HP Fabric Clustering System for InfiniBand™ Interconnect Performance on HP-UX 11iv2

Confidential Page 11 1/28/2005

Transm it Bandw idth on rx4640

(Msg. Size - 64 KB, Processor Speed: 1.5GHz, 3 CPUs.)

700

710

720

730

740

750

760

123

Num ber of Parallel Stream s

Transmit Side Bandwidth (MB/s)

0.5

1.5

2.5

Transmit Side CPU Utilization (%)

Transmit Bandw idth

Series2

Non-Optimal Configurations

HCAs when plugged in single rope PCI-X slots on a rx2600 (2 CPUs/1.5 GHz/4GB RAM) in a point-to-point

configuration offer 1-way latency of 7.8usec, as against 7.0usec latency on dual rope slots in a similar configuration.

HP-UX clustering interconnect solution saturates the single rope PCI-X slot quite fast and can only offer a bandwidth of

452 MB/s for a 4MB message, as against 760MB/s bandwidth on dual ropes slots in a similar configuration.

HP Integrity server rx4640 has shared slots. Shared slots normally operate at 66MHz and can be lowered to 33MHz

if a 33MHz card is plugged into the other shared slot on the server. Thus usage of shared slots in high performance

oriented HP-UX Fabric Clustering System interconnect solution configurations is not recommended.

Summary

HP-UX Fabric Clustering System interconnect solution offers 4.6usec 1-way latency and a single stream bandwidth of

760MB/s, all using industry standard technologies and off-the-shelf components. The solution scales well across

multiple HCAs as well as multiple connections enabling an overall improvement for real world applications. The HP-

UX Fabric Clustering System interconnect solution supports both point-to-point and switch configurations without any

negative impact on the application performance.

References

Refer to whitepapers on HP Fabric Clustering System found at www.hp.com.