Understanding Spool File Initialization Performance
The spool file initialization changed in VM/ESA Version 1 Release 1.0 to be much different and faster than VM/XA. Spool file initialization in VM/ESA is also drastically different than in HPO. (Note that these changes are also true for z/VM.) While it is much faster in VM/ESA and z/VM than in VM/XA we still are often asked "it takes me xx minutes for the spool file system to initialize on my system is this normal?". Initialization times of 10 to 20 minutes is likely for systems with 50,000 to 100,000 or so spool files. Some numbers reported by customers:
Number of Files Time for Initialization 27,833 2.7 minutes 60,000 10-15 minutes 89,000 12 minutes 200,000 + about 30 minutes
The answer is It Depends. And the following are factors on which it depends.
- Number of spool files. The more files the longer it will take.
- Number of spool volumes. Initialization spins off tasks to read SPMBKs based on the number of volumes. So along with number of volumes would be distribution of spool files across those volumes. Since load balancing in the paging subsystem is used to place files on spool volumes, the best performer tends to have the most in use space. In general, the more volumes the better the performance.
- Size of files is not really a factor unless you have very large files (perhaps large dumps and the like).
- I/O configuration. How much parallelism can we get from multiple channels or control units. Having 10 volumes all behind only 1 channel will not allow for 10 I/Os to be handled in parallel.
- Cache in CU in real-situations is most likely not a benefit (it is very helpful in normal spool activity, but not initialization). So do not expect I/O to be done at cache speed.
- Processor. At this point in initialization we are still only running on one processor, so the number of processors is not a factor. Engine speed only has small impact since this is an I/O bound process.
- Warm Start area. The data required from here is read in all at once prior to going after the spool volume data. Therefore location is not important in spool initialization performance.