There are complexities when assigning performance values to public cloud computing services that do not apply with physical servers. For one, vendor marketing statements can be difficult to reconcile with published benchmarks.
Image Source: Cloud Performance Benchmark Series; Amazon EC2 CPU Speed Benchmarks; Appendix: Comparison with Rackspace.com Computing by Kalpit Sarda Sumit Sanghrajka
On its website Rackspace makes the following statement:
Cloud Server host machines have dual quad core processors. For Linux distributions, each Cloud Server is assigned four virtual cores and the amount of CPU cycles allocated to these cores is weighted based on the size of the Cloud Server. For example, a 4 GB Cloud Server will have twice the weight of a 2 GB Cloud Server.
And on the same site:
For Windows distributions, each Cloud Server is assigned a number of virtual cores based on the size of the Cloud Server. A 1 GB Cloud Server receives 1 virtual core, 2 GB and 4 GB Cloud Servers receive 2 virtual cores and 8 GB and 15.5 GB Cloud Servers receive 4 virtual cores. Each of these cores is given equal weight when allocating CPU cycles.
However, it is hard to harmonize the above assertions with benchmark results. For instance, a group at Stony Brook found that, in most cases, the compute performance of Rackspace does not vary with instance size (image above).
If that was the only report on the issue then it may be possible to consider it a fluke, but indeed, with a few Google searches it is easy to find more evidence on the CloudHarmony blog and here.
Rackspace gives a lot more compute performance than Amazon in the smallest instance, delivering 2-7x the performance of Amazon’s "Small" instance, but with only a small amount of memory – 256 MB. The prices of the Rackspace instances double with each doubling of memory – 256 MB cost $0.015/hr, 512 MB cost $0.03/hr and so on, but the performance remains flat. After taking those two facts into account, Rackspace is most likely charging for memory, and giving the consumer a set amount of compute performance no matter what the instance size. However, it is also possible that the two statements from the website represent a minimum and that Rackspace is giving customers more compute capacity when it is available (e.g. there are unused servers). Given the disarray of benchmarks and the still young cloud market, where vendors themselves might not be able to express performance to consumers (and clock rate never worked well in that regard), it takes painstaking research to navigate through all the issues. Making simplifying assumptions in the current environment, e.g. that performance is constant linear multiple of virtual cores, is at best risky.
Comments