CPU Usage in Very Large Coherence Clusters
By jpurdy on May 31, 2012
When sizing Coherence installations, one of the complicating factors is that these installations (by their very nature) tend to be application-specific, with some being large, memory-intensive caches, with others acting as I/O-intensive transaction-processing platforms, and still others performing CPU-intensive calculations across the data grid.
Regardless of the primary resource requirements, Coherence sizing calculations are inherently empirical, in that there are so many permutations that a simple spreadsheet approach to sizing is rarely optimal (though it can provide a good starting estimate). So we typically recommend measuring actual resource usage (primarily CPU cycles, network bandwidth and memory) at a given load, and then extrapolating from those measurements. Of course there may be multiple types of load, and these may have varying degrees of correlation -- for example, an increased request rate may drive up the number of objects "pinned" in memory at any point, but the increase may be less than linear if those objects are naturally shared by concurrent requests. But for most reasonably-designed applications, a linear resource model will be reasonably accurate for most levels of scale.
However, at extreme scale, sizing becomes a bit more complicated as certain cluster management operations -- while very infrequent -- become increasingly critical. This is because certain operations do not naturally tend to scale out. In a small cluster, sizing is primarily driven by the request rate, required cache size, or other application-driven metrics. In larger clusters (e.g. those with hundreds of cluster members), certain infrastructure tasks become intensive, in particular those related to members joining and leaving the cluster, such as introducing new cluster members to the rest of the cluster, or publishing the location of partitions during rebalancing. These tasks have a strong tendency to require all updates to be routed via a single member for the sake of cluster stability and data integrity. Fortunately that member is dynamically assigned in Coherence, so it is not a single point of failure, but it may still become a single point of bottleneck (until the cluster finishes its reconfiguration, at which point this member will have a similar load to the rest of the members).
The most common cause of scaling issues in large clusters is disabling multicast (by configuring well-known addresses, aka WKA). This obviously impacts network usage, but it also has a large impact on CPU usage, primarily since the senior member must directly communicate certain messages with every other cluster member, and this communication requires significant CPU time. In particular, the need to notify the rest of the cluster about membership changes and corresponding partition reassignments adds stress to the senior member. Given that portions of the network stack may tend to be single-threaded (both in Coherence and the underlying OS), this may be even more problematic on servers with poor single-threaded performance.
As a result of this, some extremely large clusters may be configured with a smaller number of partitions than ideal. This results in the size of each partition being increased. When a cache server fails, the other servers will use their fractional backups to recover the state of that server (and take over responsibility for their backed-up portion of that state). The finest granularity of this recovery is a single partition, and the single service thread can not accept new requests during this recovery. Ordinarily, recovery is practically instantaneous (it is roughly equivalent to the time required to iterate over a set of backup backing map entries and move them to the primary backing map in the same JVM). But certain factors can increase this duration drastically (to several seconds): large partitions, sufficiently slow single-threaded CPU performance, many or expensive indexes to rebuild, etc. The solution of course is to mitigate each of those factors but in many cases this may be challenging.
Larger clusters also lead to the temptation to place more load on the available hardware resources, spreading CPU resources thin. As an example, while we've long been aware of how garbage collection can cause significant pauses, it usually isn't viewed as a major consumer of CPU (in terms of overall system throughput). Typically, the use of a concurrent collector allows greater responsiveness by minimizing pause times, at the cost of reducing system throughput. However, at a recent engagement, we were forced to turn off the concurrent collector and use a traditional parallel "stop the world" collector to reduce CPU usage to an acceptable level.
In summary, there are some less obvious factors that may result in excessive CPU consumption in a larger cluster, so it is even more critical to test at full scale, even though allocating sufficient hardware may often be much more difficult for these large clusters.