Using Multiple Protocols with Test Programs

Switching between CCITCP protocol and CCISMEM protocol can affect test program timings.

Access to the driver levels usually results in the context switching out of the sending process quite soon after the send attempt has begun. Typically the peer process uses the CPU before the context switches back to the original process.

As a result, with CCITCP, more processing time is required at the driver level than with CCISMEM which has no driver application to call. When CCITCP moves to the drivers to execute the extra data moves, the context is switched out the parent process. On the other hand, using CCISMEM with the same send operation, no context switch occurs. The CCI caller manages when to switch to its peers.

When the configuration is set to async mode (the default) File Share polls CCI to determine if data is present to read. As a result, much more time in CPU processing is spent at the user level, waiting for the next buffer to process. The application and File Share do not context switch to the conversation peer until the end of a full time-slice. That is, the CPU cannot process the next chunk of data and to respond to the original sender until the end of a full time-slice.

This can often cause quite dramatic effects on heavily loaded single CPU systems, or dual/quad core systems, when using test application created with polling loops that have scrolling screen displays.

Highlighting windows and de-selecting them alters the time spent in clients and servers because it changes their system priorities and hence the amount of time that is spent polling around not yielding to let partner processes do their work.

In these cases the fastest route is always to not high-light either client or server, then they both get the same amount of CPU with the smallest time slice. Its a case of the processes running with too much CPU not too little.

The fastest design is of course to not loop at all, then you'll be in the right place to receive the in-bound data when it arrives, and to respond quickly without wasting any CPU cycles in redundant polling.

File share can be configured to avoid the polling loop.