Performance considerations for the IEDN-enabled HiperSockets function

For network traffic within a central processor complex (CPC), HiperSockets™ processing can provide a fast response time (low latency). This fast response time is most likely for interactive (request and response) workloads where low latency is critical. However, when large amounts of data are transmitted (typically, streaming or bulk data workloads) as a significant portion of the intra-CPC workload, you might prefer OSA-Express processing due to differences in the OSA-Express processing for large data transmissions. The savings offered by OSA-Express processing for large data transfers vary; HiperSockets transmissions generally use more processor cycles than OSA transmissions, and this difference in processing cycles is magnified as transmission size grows.

If you prefer to use OSA-Express processing for large transmissions, the Communications Server IEDN-enabled HiperSockets function uses the strengths of both the HiperSockets and OSA-Express technologies, providing a configuration option that transparently controls traffic selection. Specifying NOLARGEDATA on the AUTOIQDX parameter of the GLOBALCONFIG statement transparently directs large data transmissions over OSA-Express for zBX (OSX) interfaces while directing small data transmissions over Internal Queued Direct I/O extensions (IQDX) interfaces; you can enable this setting to get the fast HiperSockets response time for small data transmissions, while avoiding the processor cost penalty sometimes incurred with HiperSockets for large data transmissions. If you do not anticipate a significant amount of large data transmissions or you always need high throughput levels (regardless of processor cycle consumption), you should not specify NOLARGEDATA on the AUTOIQDX parameter; if you do not specify NOLARGEDATA on the AUTOIQDX parameter, IQDX interfaces are favored for all transmission sizes.

Guidelines:

When NOLARGEDATA is specified for the AUTOIQDX parameter, Communications Server forces large TCP transmissions to the OSX path. Communications Server defines a large transmission as a TCP socket send of 32 K or larger. All other traffic is transmitted over IQDX interfaces.
When an OSA adapter is shared among two or more logical partitions (LPARs), the processor savings offered by the OSA adapter is less than the savings offered by configurations that use multiple OSA adapters that are not shared. Some OSA assist functions, such as segmentation offload, are not available for the shared OSA configuration.
Regardless of whether NOLARGEDATA is specified for the AUTOIQDX parameter, if large data workloads are part of the overall IEDN workload, then you should use the HiperSockets multiple write assist with IBM® System z® Integrated Information Processor (zIIP) function, which minimizes the increase in general-purpose processor usage that is incurred with HiperSockets. You should also use the 64 K Internal Queued Direct I/O (IQD) frame size.
To be sure to take advantage of IQDX for streaming workloads, use an IQDX frame size larger than 16K if you are using an MTU size larger than 8192 bytes for OSX. Otherwise, certain streaming workloads are not able to use IQDX.