Disk IO and throughput benchmarks on Amazon's EC2 | Hacker News Reader

Disk IO and throughput benchmarks on Amazon's EC2

stu.mp

37 pointssstrudeau14 comments

EC2 isn't the fastest, I'll admit. What makes EC2 a great choice for a hosting platform is that it is fairly consistent, stable, well-documented, and allows you to experiment much more rapidly than really any other option. If you go with dedicated hosting, you better hope you set everything up properly with the right balance of CPU/Memory/Disk the first time, or you've got to make expensive, painful changes to your infrastructure.

EC2 is to the hosting world as McDonald's is to the restaurant world. Consistent, predictable, obvious.

But also one of the most expensive... Doesn't really fit that analogy.

If it's cheap or not depends on what you compare to. Most other VPS systems are like home cooking: it's cheaper (and likely tastes better) but you have to do more work yourself. EC2 is more like McD: the price is higher but all you have to do is eat the thing / run the actual app. You can also buy just a cheese burger when you need it, you don't have to order a big meal.

I agree for the most part, but I would not say EC2 is consistent. Maybe if you are light enough that you don't push its limits, but if you do, you see spikes. This and the fact that nodes frequently (at the scale I spin up) don't fully boot and sometimes (rarely, though) disappear.

I'm not complaining. EC2 has been incredibly valuable and I could tack on many positive adjectives - consistent just would not be one of them for me.

It's worth noting that you can speed up write operations on EBS volumes by first initializing them by writing data to every block on the volume. Subsequent writes will be faster due to the way Amazon virtualizes disks.

http://docs.amazonwebservices.com/AWSEC2/2007-08-29/Develope...

Those docs refer to ephemeral storage. I haven't seen any claims by Amazon or test results that indicate it is also true for EBS volumes. I have been meaning to test this myself.

  gracie:~# time dd if=/dev/zero of=/dev/sdl bs=1024k count=10240
  10240+0 records in
  10240+0 records out
  10737418240 bytes (11 GB) copied, 344.745 seconds, 31.1 MB/s

  real    5m44.748s
  user    0m0.025s
  sys     0m10.152s

  gracie:~# time dd if=/dev/zero of=/dev/sdl bs=1024k count=10240
  10240+0 records in
  10240+0 records out
  10737418240 bytes (11 GB) copied, 217.106 seconds, 49.5 MB/s

  real    3m37.125s
  user    0m0.023s
  sys     0m10.268s

  gracie:~# time dd if=/dev/zero of=/dev/sdl bs=1024k   count=10240
  10240+0 records in
  10240+0 records out
  10737418240 bytes (11 GB) copied, 247.319 seconds, 43.4 MB/s

  real    4m7.323s
  user    0m0.036s
  sys     0m9.794s

EDIT: I ran it a third time just to make sure the results held. I'd still take this with a grain of salt, in my experience EBS i/o results are pretty inconsistent.

data rules

Good article and quantified my own take, making me feel justified for setting up RAID0. Lately, I have spent 5 to 10 hours a week doing setup/experiments on EC2 + EBS (about half on customer projects, and about half for my own stuff). I expect (hope!) that this time spent per week will decrease drastically since although I have been having a lot of fun, this is time spent not doing development.

I continue to be a big fan of Heroku because it saves me so much time, but I have been having limited success getting customers to go with Heroku - they seem to want to pay for custom EC2 setups.

BTW, one huge advantage of running on EC2 is that it is so easy to run an old version of a system to test against. I moved my own reserved instance from booting on an ephemeral drive (out of S3) to booting off of EBS. I only deployed what I currently need, and will always have the old version around to test against, check a bootup script, etc.

Another thing to consider is that EBS offers snapshot backups and is therefore not functionally equivalent to the ephemeral disk which is not durable.

Ephemeral disks offer more predictable performance, especially on the largest instances, however they will not provide the random iops of a raid set of EBS volumes.

We use EBS for all of reddit's databases. We occasionally run into IO limits when doing a vacuum, but that is about it.

Do you RAID them? What size instances do you use? EBS seems susceptible to poor performance when there is contention on your instance's shared network card or contention on the EBS device itself.

Vacuums use sequential IO, so it makes sense that you max out pretty quickly.

> Do you RAID them?

> What size instances do you use?

Databases are m1.xlarge.

Reply on news.ycombinator.com