ECM user scaling for Spectrum Scale NSD

Edit online

In this topic, the average ICN transaction response times are compared for the ECM cluster for both Spectrum Scale cluster configurations SD and NSD.

In this measurement series, for the Spectrum Scale virtual network (z/VM VSWITCH):
  • The default MTU (1492 bytes) was used for SD.
  • Jumbo frames (8192 bytes) were used for NSD.
For reference purposes, the response times for the single ECM node with XFS were also listed.

In Figure 1, the hypervisor CPU (total z/VM CPU load) for the entire SUT are shown as bars for both Spectrum Scale cluster configurations. The cluster configurations were compared for all SUT load levels, starting from low (8 IFL processors used), up to high (24 IFL processors used).

Figure 1. Spectrum Scale cluster configurations SD and NSD

This graphic provides an overview of how the CPU loads and response time vary with Spectrum Scale cluster configurations SD and NSD

Observations

The average ICN transaction response times are shown in Figure 1 as red, orange and blue dashed lines.
  • For both Spectrum Scale cluster configurations (SD and NSD), the response times were almost equal from the low to the medium-high load levels.
  • For the highest load level, the response time was only slightly higher for NSD but still remained around 300 ms like for SD. However, this was still significantly lower than for the single ECM node with XFS that showed a response time around 500 ms.
The bars in Figure 1 for the total z/VM hypervisor load were almost the same for both Spectrum Scale cluster configurations. At the high load level, the Spectrum Scale NSD setup consumed around 0.6 IFL processors more than SD.
Note: Compared to SD, the NSD cluster configuration also had three additional virtual servers and performed additional network block I/O.

Conclusion

For the ECM SUT, Spectrum Scale worked excellently together with the four-node ECM cluster.
  • The Spectrum Scale cluster configurations SD and NSD performed at almost the same level when network block I/O tuning was applied for NSD.
  • Especially with z/VM virtual networking, the use of network jumbo frames is beneficial for the NSD setup (see also Spectrum Scale network block I/O).

The additional CPU costs for NSD were moderate (less than 1 IFL in our setup).

The best overall performance was provided by the four-node ECM cluster with Spectrum Scale Shared Disk (SD). At least for this relatively small cluster, Spectrum Scale SD required less resources (since there are no additional NSD servers) and was less complicated to set up.

Once the overhead and manageability for directly attaching a SAN to an increasing number of cluster members becomes significant, the NSD model is a good alternative.