Advanced Computing in the Age of AI | Tuesday, March 19, 2024

Rackspace Flips The On Switch For OnMetal Instances 

A month ago, Rackspace Hosting previewed its OnMetal bare metal servers, which are configured quickly and sold with utility-style pricing like its virtual machine instances. Now Rackspace is ready to sell OnMetal capacity and believes it can offer a compelling performance and price/performance advantage compared to other clouds that run atop hypervisors.

As EnterpriseTech explained when the OnMetal service was introduced back in June, there are some workloads that simply do not perform well in a virtualized environment. NoSQL databases, for one, are particularly sensitive to latency, as Rackspace found out with its own MongoDB service and as news aggregator and early Hewlett-Packard Moonshot adopter InkaBinka explained to us recently when it talked about moving its workload, which is back-ended by Couchbase. On virtualized machines running out on the public cloud, the unpredictability of latencies between nodes is an issue and leads to overall unpredictability of the overall application. There are many other applications – particularly distributed analytics and simulation applications – that run across many nodes and that are often tuned to a specific hardware configuration to get maximum performance and, importantly, predictable performance because there is no other workload running virtually on the machine that can adversely affect these workloads.

To get around this issue, Google uses more lightweight containers to encapsulate workloads and make them behave well alongside each other on a single server node, and a number of cloud providers are getting in line behind the Docker container and application packaging system and will no doubt offer similar capabilities as this technology matures. Rackspace may yet someday offer cloudy slices based on Docker or raw LXC Linux containers, but for now, it is bare metal or a full-on XenServer virtual machine. Of the major public clouds, SoftLayer, now part of IBM, was the first to offer bare metal servers and has a big piece of its business driven by this service. Google Compute Engine, Microsoft Azure, and Amazon Web Services do not offer bare metal server slices on their public clouds. Yet.

The OnMetal service runs atop modified versions of the three-node server enclosures that were originally created by Facebook and which have been tweaked by Rackspace for its own use. The service uses the Ironic bare metal plug-in for the OpenStack cloud controller to provision the iron, and Rackspace has done a lot of work with the OpenStack community to get this tool up to speed.

The OnMetal service comes with three different server configurations, and their pricing is shown below:

rackspace-onmetal-pricing

In a blog posted by Ev Kontsevoy, the Rackspace product manager behind the OnMetal service, the pricing for the three different configurations was divulged for the first time. As it turns out, Rackspace is going to bill for OnMetal slices on a per-minute basis, not monthly or hourly, although the pricing above shows the monthly billing for each bare metal server type based on an average of 438,000 minutes per month. And to be further helpful, Kontsevoy went into the Cloud Servers online discount calculator and estimated what the street price for each OnMetal instance might be. Rackspace has volume discounts built into its pricing, and they go deeper if you prepay for capacity and deeper still as you commit to longer term usage of the capacity. The discounts start as low as 2 percent for a six-month commitment with no prepayment to 37 percent off with a 36 month commitment with prepayment. The discounts shown are roughly a third off the list price.

As promised, Rackspace is offering three different OnMetal configurations, and obviously it is not possible to offer the wide variety of core, memory, and I/O options that is possible with virtualized infrastructure. These are named onmetal-compute1, onmetal-memory1, and onmetal-io1. The I/O configuration of the OnMetal service is actually pretty beefy on the compute and memory as well as having a pair of LSI Nytro WarpDrive PCI-Express flash cards to give it very high I/O operations per second. To be specific, with 256 KB files, this flash card has a read bandwidth of 2 GB/sec and a write bandwidth of 1 GB/sec. The read IOPS using 8 KB files is 185,000 and writes come in at 120,000. These Nytro cards have a street price of around $15,000 or so each and must represent the bulk of the hardware cost of the onmetal-io1 setups, far outweighing the expense of the processors and memory.

To give some sort of competitive analysis for the OnMetal instances, Rackspace ran a bunch of rudimentary benchmarks on the OnMetal instances. On the SysBench CPU benchmark, which gauges how long it takes to complete 10,000 instances of calculating all prime numbers from 3 to 128,000. (Yes, I know 128,000 is not a prime number.) On the SysBench test, the onmetal-compute1 instance ran the SysBench test about 15 percent faster than an AWS c3-4xlarge instance, which has sixteen virtual cores, 30 GB of memory, and two 160 GB solid state drives. (Notice how Rackspace is going for the expensive, more capacious, and higher-performing PCI-Express flash card while AWS is going with the lower capacity and presumably cheaper SSDs.)

rackspace-onmetal-benchmark

If you do the price/performance analysis on this rudimentary test (which Rackspace did not do for some mysterious reason, which is why you keep me around). At the $550 list price for the onmetal-compute1 instance, this Rackspace setup gives about 21 percent better bang for the buck than the AWS c3-4xlarge instance, which costs around $605 for a month at 84 cents per hour. If you want to be generous and compare the discounted price for Rackspace, you have to do the same for AWS.

On the UnixBench test, which is a system-level benchmark running a variety of ancient generic integer and floating point tests as well as other tests to stress for file transfers, graphics, and operating system function calls, the OnMetal bare metal compute configuration was rated nearly twice as high as the AWS c3-4xlarge instance, and it offered better than twice the bang for the buck as the AWS instance as well if you do the math.

Finally, on a PostgreSQL Pgbench benchmark test, which is a transaction processing load, the onmetal-io1 instance was able to process around 15,500 transactions per second, compared to around 5,500 transactions per second for a Performance Cloud Server 2-120 configuration. That latter virtual server setup, which runs on XenServer, has 120 GB of virtual memory, 32 virtual CPUs, a 40 GB SSD for the operating system and system software, and four 300 GB SSDs for data. It costs $3.84 per hour for the capacity not including managed infrastructure or managed operations services. That works out to $2,800 for the Performance Server setup for a month, or about 50 cents per TPS on this PostgreSQL test. The OnMetal I/O instance costs 12 cents per TPS, which is less than one quarter the price per unit of work.

EnterpriseAI