Conclusions

Based on the above results a few conclusions can be drawn:

• For a PIAF system processing large data sets not fitting in cache I/O and NFS performance as well as the amount of memory and CPU power become limiting factors. This can be helped considerably by enhancing the overall I/O rates as well as using faster CPUs with more memory.

• For small data sets which essentially fit in cache, the bottleneck lies in master-slaves communications. The solution in this case involves enhancing both the message passing efficiency and CPU power especially on the nodes where the masters are to be run.

• The larger the number of PIAF nodes, the higher are the required data rates over NFS. Using the parallel file system should also help considerably here.

• Depending on the expected usage different PIAF configurations can be used. For jobs using small data files and straightforward queries a smaller system is adequate. For large I/O limited jobs or jobs with complex queries a bigger configuration should be used.

• The number of concurrent users in the system is mainly limited by the available memory. In real use PIAF process sizes can grow all the way to the n-tuple cache size which is of the same order as the available memory on one CS-2 node. Although the low system load associated with interactive work would allow more concurrent jobs, increased paging due to small memory limits the performance.