The L3 HIPPI device driver runs in kernel space, while the library, T-server and IPH runs in user space. In theory further improvements could have been expected by implementing the transport layer in kernel space.
The maximum data transfer rate achieved by this implementation is 14 MByte/s which is 87% of the physical limitations of the test setup. In comparison, IBM has tested TCP/IP over the same HIPPI version between two RS6000s model 550 achieving 9 MByte/sec. By using the L3 test setup, a maximum transfer rate of 5.5 MByte/sec over TCP/IP was obtained. The TCP/IP protocol stack includes many functions which are omitted, but on the other hand TCP/IP runs completely in kernel space and is therefore not losing transfer speed due to context switches.
In general purpose networks most packets are small, not more than a few 100s of bytes, and in that context our comparison with TCP/IP when running the L3 HIPPI software with a packet size of 2 MByte might seem unreasonable. However, the purpose of this project was to make a highway for transfer of large data sets, and the test results clearly show that for this purpose our solution is better than the available TCP/IP implementation over HIPPI.
The result achieved in loopback on the model 590 indicates that two 590s running the L3 HIPPI software end-to-end possibly could have transferred data at around 56 MByte/sec. Again, this is far better than TCP/IP which has been tested by IBM to run at 38 MByte/sec between two 590s using the HIPPI version 3.2.2. We confirmed this data rate over TCP/IP by getting 19 MByte/sec when running the model 590 in loopback.
The papers "Implementation of a lightweight transport service over HIPPI on the IBM RISC System/6000" and "A proposal to the L3 HIPPI Network Software Architecture" are both available in one Postscript file (314 KB).
For more information contact Gudrun R. Dalgeir, Tel: +41 22 767 7557, E-mail: Dalgeir@cern.ch.