New Amazon EC2 Instance Type - The Cluster Compute Instance
aws.typepad.com
aws.typepad.com
I think the most noteworthy feature is the 10G Nic which is more conducive to clustering than the standard GigE Nic for other instance sizes. I'm not aware of any other public IaaS provider offering 10G network currently.
In terms of raw performance the x5570 will probably perform very well relative to other IaaS providers. EC2 is one of the few IaaS clouds that utilizes a heterogenous hardware environment to enable better scalability between instance sizes. Other providers like Rackspace Cloud use homogenous environments (Opteron 2374 in the case of Rackspace) and rely on Hypervisor CPU throttling to enable CPU scaling between different instance sizes. On the low end homogenous clouds tend to offer better value but are not well suited for HPC application on the high end (e.g. a 2GB Rackspace Cloud server kills an EC2 m1.small, but the largest 16GB instance lags far behind EC2 m2.* instances). In fact in some homogenous clouds CPU performance is fairly constant from small to larger sized instances (you are essentially just paying for more memory).
I recently wrote a blog post comparing CPU performance of 20 different IaaS providers including EC2 and Rackspace Cloud. In it I used a compilation of 19 different CPU related benchmarks to produce an ECU metric for about 150 different cloud server configurations:
http://blog.cloudharmony.com/2010/05/what-is-ecu-cpu-benchma...
The NIC is not the most interesting part, but the network is: not only do these new instances offer 10 GigE, but they also offer full bisection bandwidth to all instances in a placement group (=> no oversubscription of core switches). That is pretty unusual, and a big win for network-intensive applications.
Aha, we finally have another EC2 Compute Unit comparison than the vague "1.0-1.2 GHz 2007 Opteron or 2007 Xeon processor."
The X5570 is basically a server grade 2.93GHz Core i7 and it equates to 16.75 ECU (since the "pair" is 33.5 ECU). You can then do very vague and imprecise extrapolation and assume a regular 2.66-2.8GHz i7 is in the 12-15 ECU area (which makes most EC2 instances look rather anemic and probably explains why the Redis benchmarks on EC2 seem so poor.)
I used to use EC2, but I've found I get significantly better performance (and of course support) at Rackspace for the same price. Also avoid Rackspace cloud sites, performance is pretty awful.
Anyway back to the dedicated server point. If you use virtual servers you can very easily replicate images, spin up copies and scale your computer power on the fly. Case in point, I was running one of my apps on a $10 pm instance on Rackspace. We got a Techcrunch article out of nowhere and the site instantly died under the traffic. I resized the server on the fly and boom, * straight back up. Once the peak died down I scaled the server back down. A couple of months down the track we get near Techcrunch traffic every day and I haven't had to scale the code base at all - I just keep it running on a larger instance. Much, much more cost effective.
* ok not straight. Unfortunately when resizing a Rackspace cloud instance under heavy load it takes ages to run (30+ min isn't unusual compared with 5 min under light load). I've found that the best solution is to spin up a copy on another instance and change the dns to point to the new one. I wish Rackspace would add some way to temporarily block traffic to a server to it can scale. Some kind of "panic", scale me up, button would be great.
> This result places us at position 146 on the Top500 list of supercomputers.
And $1408 per hour.
And for almost 60k per hour, you will be number 1! How long till AWS as the capacity to top the chart?