gh@swdev.Waterloo.NCR.COM (Gord Harris) (03/24/90)
I am measuring OS/2 system overheads in an attempt to predict the CPU time available for a process control application, and consequently the appropriate PC and clock speed for a given assured throughput. Although limited tests will be performed shortly, some non-trivial combinations of software might degrade performance unexpectedly. If you have hard numbers or anecdotal information related to topics below or others omitted, please reply. A summary will be issued in two weeks or so. The measurements should assume sufficient memory to inhibit swapping and should specify processor type and frequency. Please exclude sources of bias such as CRT activity, or account for them by stating a level of confidence. Of interest are basic thread and process switching times, and whether these times remain constant regardless of the number active. The use of semaphores has been described as costly (hence new fast semaphores), but how costly? The OS/2 driver access time may prove to be similar to MSDOS but what is the typical code increment and cost to use device helper routines in mixed mode drivers? For example, the disk driver may require several milliseconds of extra setup time before activating the physical unit. Printer spooling is an anticipated background activity whose performance is only limited by the 19.2 Kbaud serial link. At this rate, how much CPU time is available for a low-priority foreground process? Conversely, can a fore- ground process be guaranteed at least 60% of the processor over a 100 milli- second interval (2 out of 3 ticks) if its priority is raised to time-critical? Lastly, is there danger to process control applications caused by threads which despite their higher priority do not run immediately after becoming unblocked? Thanks in advance for any eye-openers and the sweat getting them. -- G.Harris@Waterloo.NCR.COM uunet!ncrlnk!ncrwat!swdev!gh 580 Weber St. N., Waterloo, ON. N2J 4G5 Tel. 519-884-1710