Scale-out study results

Edit online

For the measurement series discussed in this topic, all of the described setup and tuning was applied to the SUT. To summarize, in this topic the measurement series scaled the number of virtual ICN users from 250 to 1000 for the four-node ECM WAS cluster setup.

This topic uses the ECM workload (with its different SUT load levels) that was described in ECM workload.

One of the major objectives for the scale-out case study was the performance of a clustered multiple ECM node environment (that used WAS and Spectrum Scale clusters) versus a single ECM node setup. For this purpose, the ECM clusters were scaled-out from a single node to a four-node cluster during this study. Spectrum Scale as parallel filesystem played a major role in this type of clustered multiple ECM node setup.

The ICN transaction throughput depended upon the number of virtual ICN users and the think time between transactions. The ECM cluster had four nodes with 8 virtual CPUs per node (that is, 32 virtual CPUs for the entire ECM cluster).

Note: For the scale-out case study, the average ICN transaction response time over all ICN transactions was considered as performance indicator.

The ratio of CPU overcommitment was:

1:1.6 (30 IFL processors versus 48 virtual CPUs) with Spectrum Scale SD.
1:1.8 (30 IFL processors versus 54 virtual CPUs) with Spectrum Scale NSD.

Note that the SUT z/VM hypervisor had 30 IFL processors in total.

Table 1. SUT load level definitions
SUT load level	# of virtual ICN users	z/VM total load (IFL processors used)
low	250	< 10 IFLs (0-33%)
medium	500	10 - 15 IFLs (33-50%)
medium-to-high	750	15-20 IFLs (50-66%)
high	1000	> 20 IFLs (66-100%) - see Note below

Note: The highest z/VM IFL processor usage was 80% (24 out of 30 IFL processors used) for the SUT with 1000 virtual ICN users and Spectrum Scale NSD.