Video Screencast Help
Symantec to Separate Into Two Focused, Industry-Leading Technology Companies. Learn more.
Storage & Clustering Community Blog

Cluster Asymmetry Part 2

Created: 13 Jan 2012 • Updated: 11 Jun 2014
ScottK's picture
0 0 Votes
Login to vote

The last post looked at frequency of  cluster asymmetry, and what types of asymmetry (memory, CPU, etc.) are more common. To quickly recap, a third of clusters sampled showed some kind of asymmetry. Amount of memory and server models were the most common type of asymmetry. CPU speed & CPU model were the least common.

Going to the next level of detail: how asymmetric are these clusters? For example, when two cluster nodes have different amounts of memory, is the difference in memory a lot, or a little?  Server Model & CPU Model differences are hard to quantify cardinally or even ordinally, so let’s look at Memory, CPU speed, and the number of CPUs.

Each cluster in the sample is different, since they come from different customers. Using the absolute number of CPUs, or GB of memory, or MHz or speed, would be intuitive but misleading. Consider a one CPU difference within two different 2-node clusters. Cluster A (64, 65) and Cluster B (1,2) both differ by one CPU. But the relative asymmetry is small for Cluster A  and large for Cluster B. To compare across clusters, I normalized by calculating the mean GB of memory, MHz of speed, and number of CPUs for each cluster. I then calculated each cluster’s intra-node difference from its own mean. And finally took the average difference across clusters.

Think of it this way: if the relative asymmetry is 50%, then one node in the cluster had 50% more CPUs (or memory, etc.) than the other node. The results:

The small horizontal bar is the average (mean) while the thin vertical line marks +/- one standard deviation.

Again, these are just those clusters which displayed a given type of asymmetry. So, for clusters with differing numbers of CPU across nodes, the average difference was 75%. Recalling the previous blog post, about 18% of clusters had different numbers of CPUs across nodes. Taken together, I’d say there are moderate number of clusters out there that display material differences in the number of CPUs among their nodes.

Similarly, Memory was the most frequent attribute showing asymmetry (previous post), and when it differs, it differs quite a bit (graph above).

In contrast, consider differences in CPU speeds. Few clusters showed differences in CPU speed among nodes (again, previous post), and when they did, the differences were small (graph above).