Throughput optimization with AutoDAB

Two measurements with the following access profiles were conducted to demonstrate the effects of AutoDAB:

Measurement 1:	8 tasks accessing with write access to one file each. 4 tasks accessing with a sequential access pattern, 4 with a random access pattern. There are no write I/Os.
Measurement 2:	8 tasks accessing with random access pattern and 25% write share.

The measurements were taken on an S server. File size was 500 MB and the files were located on two NK2 volumes. I/O length was 4 KB and the duration of measurement was 5 minutes each. Measurements were performed without and with AutoDAB cache. AutoDAB cache in CACHING-MODE=*READ has a cache size of 2,000 MB in the main memory.

The measurement results are shown in the following table:

Measurement	Access type	IO/s without DAB cache	MB/s without DAB cache	IO/s with DAB cache	MB/s with DAB cache
1	sequential	32,809	128	58,106	226
1	random	32,935	128	78,832	307
2	random	31,018	121	79,746	311

With AutoDAB the I/O throughput has been considerably improved. This is also true for the overall throughput.

The reasons for this are:

• The better the hit rate in a DAB cache, the more CPU power can be converted into I/O throughput.

• In measurement 1 the sequentially processed files have a very high hit rate due to the high prefetching factor, i.e. data areas are cached in advance, and are kept practically resident in the DAB cache due to the low data volume.

• Measurement 1 also shows that DAB’s effective use of the cache memory makes an even greater portion available for the caching of files with random access patterns; this also helps to increase the throughput.

• The random files in measurement 2 also have very high hit rates due to the intelligent algorithms of the AutoDAB.