* Rutherford Research to examine jumbo packets
Last time, I discussed how jumbo frames may be useful in grid or server cluster environments. Some work on such usage is being carried out by Canada’s Rutherford Research.
Rutherford Research, which offers professional services for research and development, is conducting its tests across CA*net 4, a high-performance optical network spanning Canada and funded by its government.
The company notes that using jumbo packets can double performance over Gigabit Ethernet in high-performance computing environments, and can more than double performance in networks running at even higher rates.
“While the use of jumbo packets might seem a wizard’s tuning trick, it represents a logical and immediately valuable benefit to grid computing,” the company writes on its Web site.
But how much of a benefit? The company says it is looking at how big of a performance gain one can get from using jumbo frames with various distributed applications, including massive data transfer (the most obvious application), collaborative visualization and distributed file systems.
The company’s initial test bed conducts TCP transport across Gigabit Ethernet sites. Using SGI servers and Linux and SGI workstations, the firsts tests showed throughput of 400M bit/sec “with significant potential for improvement.”
Next, the company plans to test the effects of packet size changes on NFS Version 4. It says that in NFS end node fragmentation and reassembly can really hurt performance, and is looking at increasing the maximum transfer unit to 9,000 bytes. The company says it believes it can get throughput that is vastly improved, close to that of a local storage-area network using Fibre Channel.
