Benchmarking Storage Systems, Part 1 Page 4


Want the latest storage insights?

Download the authoritative guide: Enterprise Data Storage 2018: Optimizing Your Storage Infrastructure

Fibre Channel Switches

The question is simple. Do your applications require full duplex 2 Gb I/O performance any port to any port? Most environments do not. For the few that do, the Fibre Channel switch can be a big issue, as not all switches are alike.

NIC and TOE cards

These cards are used for IP communication and are part of a NAS benchmark. NICs (network interface cards) provide the connection between the computer and the network's physical medium, while TOE (TCP/IP offload engine) cards offload the processing of the TCP/IP stack, taking some of the burden off of the main processor.

I have seen over 2x difference in performance and almost the same for system overhead for Gigabit Ethernet NIC cards from two different vendors using the same host, same OS, same everything. Therefore it is important to get the right TOE or NIC card for the type of work that you are going to be doing.

Failover and Fail Back

You might ask why failover and fail back would be part of a benchmark and not just a usability issue, and you might be correct. The reason I have added it to this list, though, is that in some cases fail back requires manual intervention and can significantly impact performance.

Failover performance might need to be tested for both HBAs and possibly the RAID controller depending on the design. You should know the performance of the system during a failover, as HBAs and NICs generally have higher failure rates than most of the other server system components.

What's Next

We have covered the components of a storage benchmark and some of the issues surrounding these components. Each of these hardware and software pieces can have a significant impact on the performance, reliability, and price of the system. Regardless of the type of NAS or SAN system, storage vendors will either provide each of these parts or you will be using some you already have. All of the parts must be considered and specified as a possible function of the benchmark to ensure that you need – as compared to get – what you asked for.

Future articles in this series will cover how to develop representations of your workload and then how to package it all together for a formal benchmark.

» See All Articles by Columnist Henry Newman

Submit a Comment


People are discussing this article with 0 comment(s)