Enabling DynaCache disk offload

Edit online

In these tests, the dynamic cache size was adjusted to a value that would allow the cache to offload entries from real memory to disk.

The number of workload generating clients was set to 65 and the disk offload was enabled. The eviction policy was set to random. The dynamic cache size had to be reduced to less than 2000 to obtain any observable data off-loaded. The smallest value that WebSphere® would accept for a Dynamic Cache size was 100 statements.

The environment with the WebSphere Proxy Server was used for this test. In addition to the normal workload, the modified Trade workload simulator script from the previous run was used, which had removed all transactions that resulted in a data base update for one of the runs (0% update part).

All runs were done with the cache configuration described in Table 1.
Table 1. Cache definitions - Enabling DynaCache disk offload
DynaCache size Caching mode Servlet caching Replication domain Replication type
2000, 1000, 500, 100 Distributed map Disabled Entire domain Not-shared
Figure 1 illustrates the throughput when adjusting the dynamic cache size to permit offloading of entries from memory to disk, both with and without updates. Figure 2 illustrates the CPU utilization when adjusting the dynamic cache size to permit offloading of entries from memory to disk, both with and without updates.
Figure 1. Throughput - Comparison disk offload versus no disk offload with and without updates.
This figure shows the throughput comparison of disk offload versus no disk offload with and without updates.
Figure 2. CPU utilization comparison - disk offload versus no disk offload with and without updates.
This figure show the throughput for comparison disk offload versus no disk offload with and without updates.
Figure 3 illustrates the CPU utilization results when using disk offload, expressed relative to the results using a cache size of 20,000 and not using disk offload, and taken from the perspective of: the z/VM® LPAR, the WebSphere Application Server, the DB2® server, the disk offload proxy server, and the firewalls.
Figure 3. Changes in CPU utilization for DynaCache Disk Offloading relative to base (cache size 20,000 and no-disk offload).
This figure shows the changes in CPU utilization for DynaCache for Disk Offloading relative to base (cache size 20,000 and no-disk offload)

Observations

For the 14% update (normal), Trade workload disk offload produced no improvement in throughput until the dynamic cache size was set to its smallest value (100). At this value, disk offloading increased throughput by 15%. CPU utilization for these cases was almost identical.

It was noticed that the Proxy Server CPU utilization was greater than 90% for the runs with the 14% update Trade workload. For the no update Trade workload, the Proxy Server CPU utilization increases further, according to the increased throughput, and reaches the full CPU utilization. That means, it was gating the throughput. Therefore, the number of virtual CPUs assigned to the Proxy Server guest was increased to two for the no update Trade workload tests.

For the no update Trade workload, disk offloading resulted in an immediate advantage and the advantage increases as the dynamic cache size decreases.

Figure 3 shows the advantage of increasing the number of CPUs.

An interesting observation is the relationship between the no update Trade workload case in the disk offloading tests, and the results seen in DynaCache scaling section, Figure 1. At a dynamic cache size of 100 the disk offload throughput case is only 11% less than the maximum throughput seen in the no update Trade workload update case, seen above, with a dynamic cache size of 10000.

At a dynamic cache size of 100, disk offloading improved throughput by more than 50%.

In total, there is a slight reduction of CPU utilization for the z/VM LPAR, especially in the no update case.

The CPU utilization for the proxy server and firewall increases over the base results for the no update case. This increase is related to the increased throughput obtained in the no update case.

It is interesting to notice the difference in the CPU utilization for WebSphere and DB2 for the no update case. With disk offloading set to yes and a DynaCache size of 100, there is a significant reduction in CPU utilization versus the disk offloading set to no case.

Conclusion

Disk offload can provide a significant performance improvement for the right application transaction mix and can reduce the CPU load. Even in the case where it provides no significant improvement, it provides no additional overhead. That means that in the tested environment, this feature could be used to run with a significantly smaller cache size, for example of 2000 statements, to save memory without performance degradation.
Note: Disk offloading is a very important DynaCache feature.