Review: SoftLayer's cloud is fast and flexible
SoftLayer brings fine-grained configuration options, high performance, and interesting extras to the self-service cloud
By Peter Wayner
Contributing writer, InfoWorld |
Layering on the options
The hardware is only part of the shopping process with SoftLayer. While a number of cloud providers give you just a few radio buttons of options during the configuration process, SoftLayer takes you through four pages of choices. Four extra public IP addresses are 1 cent per hour. A premium monitoring package is 6 cents per hour. You can add five different disks to your server if you like. You can add more local disk space or storage on a SAN. It's like the old days when the server salesman wanted to fill up all the bays with extras. If you want a license to run Windows, SoftLayer will toss one in for between 5 and 10 cents per hour depending upon which version of Windows you choose.
While I had no real problem configuring several machines, SoftLayer is still working through making all of this function smoothly. I asked for MongoDB on my machine but got a message later that it wouldn't work with Ubuntu 12.04, the OS I happened to choose. There are menu items on the portal for CPanel software licenses, but I wasn't given an option to buy one. SoftLayer is clearly planning on making it easy to buy extras, but not all of the dots are connected yet.
One interesting option is a "bare metal" server, also sold by the hour or by the month. I spun up one of these with two cores running at 2GHz and 2GB of RAM at a price of 50 cents per hour. These don't run in the same seemingly endless stack of virtualization, allowing them to access the I/O channels faster. This pays off with databases and other disk-bound applications.
The performance of the "bare metal" server was often better, but not in every case. In some the results were largely the same as the "painted metal," for lack of a better term. The times were about the same for the relatively linear, single-process jobs like the Batik vector graphics rendering or the Eclipse test. These are largely computational. But other tests such as the Lucene searching or the Sunflow ray tracing sped up dramatically because the code was able to take advantage of the extra cores and the better disk I/O. The DaCapo benchmarks have an option to limit the number of threads, and when I held the bare-metal machines to one thread, the gains largely disappeared.
Your results, of course, will differ just as the results from the benchmarks do. The so-called bare-metal machines are better at handling I/O operations such as writing to disk because they don't have the hypervisor adding an extra step to the interaction with the device drivers.
