Really? I know reserved instances lower the price but they still have shown themselves less performant for the dollar in the benchmarks I have seen. Are there any good articles you could recommend with benchmarks showing AWS to be more cost effective?
My experience is similar. EC2 performance is mediocre in general and quite poor compared to dedicated bare metal. Disk I/O is abysmal (yes, even on "SSD" EBS).
Any citations? There are lots of people using EC2 at appropriate instances sizes to do compute intensive loads. Provisioned IOPS and, again, appropriate instance types are necessary to get high performance disk I/O.
My "citation" is my own experience managing infrastructure for several AWS-backed startups. I think public cloud services have a role and make sense in certain cases (particularly for very early stage companies), but I'm realistic about their performance characteristics and tradeoffs. The biggest danger is in overuse of proprietary AWS services, which makes it much more difficult to migrate off when the business outgrows the public cloud.
EBS is not intended for high-perf applications. This is well-known and well-understood. As far as CPU perf being poor compared to bare metal--you might be the one person out there who actually needs it, but I'm betting not.
I use AWS, and will continue to use AWS, because it means less screwing around with stupid things and more time making important things work.
CPU-bound workloads are not particularly rare. Also, EBS is your only storage option for any data you might care about. 90% of the reason RDS exists is because I/O is so bad on normal EC2 instances.
You can run a database just fine using provisioned IOPS EBS especially in a RAID 10 configuration. I'd like to see some data to the contrary as I've actually done it.
There's some disadvantages to using RAID 0 with EBS volumes in a performance dependent situation.
Since you're writing to multiple EBS volumes for the RAID 1 portion of the RAID 10, you're going to require more EC2 to EBS bandwidth.
EBS volumes are replicated across multiple servers, so you don't necessarily need the reliability, especially if you're replicating the data elsewhere.
YMMV, of course, and everyone has different priorities and different levels of what constitutes acceptable risk, but RAID 0 with EBS isn't quite the data death wish it is with physical hardware.
> EBS is your only storage option for any data you might care about.
No, it's not. It is if you want data-at-rest storage, but I don't care about that when I have multi-master MySQL and Postgres running on instance stores. Or when I have replicative NoSQL databases (Riak is more than happy running forever in instance stores and duping across AZs).
I only have a general comment about this, but that kind of performance is almost not worth thinking about in most cases. Just like NBA players need to be "tall enough", your cloud needs to be "fast enough", and AWS is just that.
Having said that, nothing beats getting real hardware in terms of performance/$. I used to work for a company that exclusively worked on SoftLayer's hardware servers. These were $500+/month each, but they were fast. The point is that if you can devote a dual octocore machine to what you are doing, and are willing to pass up on the SAN, flexible networking, etc. then you get a very fast box. If what you are doing requires lots of very fast hardware, then yea AWS or anything like it is not really for you. But if you are like most people, your AWS bill is not going to break the bank and you can just spend one less billable hour figuring out why the server is slow, and just double its size.
Compared to provisioned IOPS SSD, it's about the same. Compared to regular "burstable" SSD it's much much better. The gp2 EBS SSD is not actually meant to be used for data storage, something which AWS hasn't done a good job of making clear.
Flexibility of networking, Regional coverage, Tooling , Performance,
On the cost side with reserved instances it's less expensive.