Intel AS/400 RISC Server manual Virtual Ethernet CPU Cost and Capacities, VE Capacity Comparisons

The blue square line shows an iSCSI connection with a single target iSCSI HBA - single initiator iSCSI HBA connection, configured to run with standard frames. The pink circle line is a single target iSCSI HBA to multiple servers and initiators running also running with standard frames. With the initiators and

switches configured to use 9k jumbo frames, a 15% to 20% increase in upper capacity is demonstrated.15

17.6 Virtual Ethernet CPU Cost and Capacities

If the virtual Ethernet connections are used for any significant LAN traffic, you need to account for additional System i CPU requirements. There is no single rule of thumb applicable to network traffics, as there are a great number of variables involved.

The charts below demonstrate approximate capacity for single TCP/IP connections, and illustrates the minimum CPW impacts for some network transaction sizes (send/receive operations) and types gathered with the Netperf 16exerciser. The CPW chart below gives CPWs per Mbit/sec for increasing transaction sizes. When the transaction size is small, the CPW requirements are greater per transaction. When the arrival rate is high enough, some consolidation of operations within the process stream can occur and increase efficiency of operations.

Several charts are presented comparing the virtual ethernet capacity and costs between an iSCSI server running jumbo frames and standard frames, and a IXS or IXA server. In addition, a comparison of costs while using external NICs is added, to place the measurements in context. The “Point-to-Point” refers to the cost between an iSCSI, IXS or IXA attached server and an host system across the point to point connection. “Port Based VE” refers to a port-based connection between two guest servers in the same partition. “VLAN based VE” refers to a Virtual LAN based connection between two guest servers in the same partition, but using the VLAN to port associated virtual adapters. In the latter two cases, the total CPW cost would be split across partitions if the communication would occur between guest servers

hosted by different partitions17.

17.6.1 VE Capacity Comparisons

In general, VE has less capacity than an external Gigabit NIC. Greater capacity with VE is possible using 9k jumbo frames than with using standard 1.5k frames. Also, the iSCSI connection has a greater capacity

15In addition, jumbo frame configuration has no effect on the CPW cost of iSCSI disk operations.

16Note that the Netperf benchmark consists of C programs which use a socket connection to read and write data between buffers. The CPW results above don’t attempt to factor out the minimal application CPU cost. That is, the CPW results above include the primitive Netperf application, socket, TCP, and Ethernet operation costs. A real user application will only have this type of processing as a percentage of the overall workload.

17Netperf TCP_STREAM measured on a System i Model 570 - 2-way 26F2 processor (7495 capacity card), rated at 6350 CPWs, V5R4 release of i5/OS. The IXA attached server was a x365xSeries (4way 2.5Ghz Xeon with IXA and Windows Server 2003 with SP1. The iSCSI servers were HS20 BladeCenter 32.Ghz uniprocessor servers with a copper iSCSI (p/n 26K6489) daughter card. Switches were Nortel L2/3 Ethernet (p/n 26K6524). This is only a rough indicator for capacity planning, actual results may differ for other hardware configurations.

IBM i 6.1 Performance Capabilities Reference - January/April/October 2008
© Copyright IBM Corp. 2008	Chapter 17 - Integrated BladeCenter and System x Performance	289