Disk I/O for ECM nodes

In this topic, the disk I/O rates for the single ECM node versus the ECM cluster are compared at different load levels.

Figure 1 shows the disk I/O rates (read and write) for the individual ECM nodes. The disk I/O rates are displayed as stacked for the individual nodes. The read/write rates are also displayed as stacked.

Therefore, for the ECM cluster there are:
  • four values for read that are added together in a stack.
  • four values for write that are added together in a stack.
Spectrum Scale SD was used for both setups, and the SAN disks were directly attached to the ECM nodes.
Figure 1. Spectrum Scale disk I/O - single ECM node vs. ECM cluster

This graphic provides an overview of how the disk I/O varies with a single node versus a multiple node cluster

Observations

  • For the low load level, the disk I/O rates (read+write) were less than 50 MiB/sec for the single ECM node, and around 100 MiB/sec for all nodes of the ECM cluster.
  • At the high load level, the disk I/O rates was around 115 MiB/sec for the single ECM node, and 225 MiB/sec for all nodes of the ECM cluster.
  • The ECM cluster had almost double the disk I/O rates compared to the single ECM node.
  • The write I/O (shown as blue bars) was almost the same level for the single ECM node as for the ECM cluster over the complete load level bandwidth.
  • The read I/O (shown as red bars) was almost 3 times higher for the ECM cluster compared to the single ECM node.

Conclusion

The amount of updated or new ECM content was similar, independently of whether a single-node or multiple-node ECM system was being used. Therefore, the write I/O rates were similar for both setups.

However, the read I/O rates were significantly higher when using an ECM cluster.
  • In a Spectrum Scale cluster with multiple nodes every node had its own pagepool, whereas there was only one unique pagepool for the single ECM node.
  • In a Spectrum Scale multiple-node cluster, the effort required to synchronize the pagepools was higher. In addition, user data had to be re-read from the NSD disks more often.
  • Therefore, the overall resulting read I/O was higher in the case of a ECM cluster.