By Lionel Gibbons | November 14, 2016 | HPC
We wanted to find out more about how our customers are deploying and using their clusters in the real world. We asked them a series of questions about how big their clusters are, how many they have, and whether they plan to expand in the future. We also asked what hardware and software our customers have deployed, and what they are using these clusters for in their day to day operations.
A number of the questions wouldn’t be interesting to most of you. However, we also asked questions that point to the current state of affairs in clustered infrastructure, and hint at what may be coming in the future
We also compared the results of this latest survey to one we did last year. Some of what we learned was surprising and I wanted to share a few of the highlights with you.
One thing that the latest survey seems to indicate is that customers are running bigger clusters. Many of them are much bigger. According to the survey, 43% of respondents are operating clusters with more than 100 servers in them. Only 10% had clusters with 10 or fewer servers. This is impressive, although we need to keep in mind that this was a survey of Bright customers, so the respondent pool consists of people who grasp the value of using cluster management software. However, even those running smaller clusters may want to start using professional management software now. They should at least consider it, if they expect their clusters to grow appreciably in the coming year. That way, they’ll be ready for anything.
The other interesting number that jumped out at me was the number of customers that reported operating multiple clusters. A few reported having more than 20 clusters, but fully two-thirds reported that they operate more than one cluster. Most (about 40%) reported having two (26%) or three (14%) clusters.
To the question asking how many clusters users operated in, “geographically separated data centers,” 32% said two, 9% said three and 15% said more than three. It seems clusters aren’t just for the local site anymore.
Curious as to what was in those clusters, we asked about workload managers and whether or not customers had accelerators installed. SchedMD’s Slurm is by far the workload manager that most Bright customers are using, with Adaptive’s TORQUE and Altair’s PBSpro closely tied for second place. Other notables include Moab, Univa Grid Engine and LSF.
When it comes to accelerators, Bight customers are into performance. Nearly three-quarters of respondents use some kind of accelerator in their clusters. That’s up from 48% in last year’s survey. Increased use of high performance data analytics and deep learning are likely contributors to this shift. We’ll be keeping an eye on this number in future surveys.
There’s more of the survey to talk about, but I’ll save that for another post. For now, my takeaway from the results of the deployment questions is that clusters are alive and well. People are running some very big clusters, and they’re putting some serious hardware in them to boost performance. Clearly, these customers consider clusters a critical part of their organization’s business.