New EC2 Instance Type - The Cluster GPU Instance
aws.typepad.com
aws.typepad.com
http://cloudscaling.com/blog/cloud-computing/amazons-ec2-gen...
And he estimated an overly conservative 10-20% annual growth. But given the EC2 buzz this year, and personal anecdotes from my friends and colleagues using it, my gut feeling tells me the 2010 revenues will have increased 50-100% over 2009 revenues.
Is EC2 profitable to Amazon? Likely very profitable if you want my opinion. It is well accepted in the industry that the dominant cost in large scale datacenters is power and cooling --not hardware, not human resources-- and I keep running numbers in my head and the hourly prices of all instance types are well above power & cooling.
Just as an example, we know that this new GPU instance has two 95W Xeon X5570 and two 247W Tesla M2050; assuming (1) a max TDP of 50W for the motherboard and rest of the server, (2) instances are run under full load 100% of the time and always reach these max TDP numbers (unlikely, but follow me for the sake of the argument), (3) Amazon uses servers with 80PLUS power supplies (80% efficient or more), (4) a rather good datacenter with a PUE of 1.3 (power usage effectiveness, which includes overhead from power distribution and cooling; numbers in the range of 1.2-1.4 are often quoted by James Hamilton from the AWS team: http://perspectives.mvdirona.com/), and (5) electricity costs of $0.10/kWh (average in the US, but I know Amazon datacenters are in locations with cheaper electricity), then the hourly power and cooling costs would be:
(95*2 + 50 + 247*2) / 0.8 * 1.3 / 1000 * 0.10 = $0.119/hr
Given these numbers, Amazon must recoup the initial deployment costs very, very quickly... Which is why I also think EC2 must be very profitable.
From James Hamilton (who just happens to work for Amazon) http://perspectives.mvdirona.com/2010/09/18/OverallDataCente...
Anyway, this does not change my point that the EC2 hourly prices are so high that they eclipse both power & cooling as well as server costs for Amazon.
As a rule of thumb, if Amazon is doing something then it is profitable. Amazon will very quickly stop doing something if it can't make money.
New to the whole Bay Area thing eh? Plenty of Zombie Startups out there that aren't getting traction/pivoting.
I don't know of any major provider that offers SSD instances. It really is an untapped market.
People, stop assuming that $/GB is the only metric that matters.
I also don't know if this is still a problem but there was a worry about limited writes on SSDs and if normal harddrive lifetimes are greater than that of an SSD by a significant amount I could see why it would be a problem.
Linode: Timing buffered disk reads: 230 MB in 3.01 seconds = 76.44 MB/sec
My local X-25M: Timing buffered disk reads: 246 MB in 3.02 seconds = 81.48 MB/sec
This would suggest they are using SSDs (or fast RAID? I'm not sure if this benchmark is sequential access). I get comparable performance for the Georgia data center, but it varies greatly there, so I'm not sure what's going on.
EDIT: Sorry, apparently I can't tell my sde from my sda. My actual SSD performance is:
Timing buffered disk reads: 564 MB in 3.01 seconds = 187.47 MB/sec
So yes, disregard everything above.
I doubt they're using SSDs.
As for your X25-M, do you have TRIM enabled? Your read speeds are consistent with a used-state SSD without TRIM.
Btw here is hdparm reading for my Western Digital Caviar Black (regular hard disk) just to give you an idea: 318 MB in 3.01 seconds = 105.56 MB/sec
Timing buffered disk reads: 142 MB in 3.00 seconds = 47.28 MB/sec
I think it's more likely they are using a high performance shared storage box. Those are really fast and reliable. The only reason I would use disks (whether SSDs or spinning metal) would be for speed. If, say, a storage box has 200 servers hanging from it and all 200 decide to go nuts with disk access you won't be able to sustain anything near 76 MBps per box.
As for the storage boxes themselves, they usually employ piles of ECC RAM, specialized network and disk controllers, SSDs, fast and small disks, larger and slower disks and keep moving data around trying to guess the optimal positioning to give you the best possible performance under your varying workload. It must be really cool to design one.
hdparm only does sequential reads, so don't expect a night-and-day difference between SSDs and rotating disks. Plus that benchmark reads out of the disk (buffer) cache, so you might really testing your memory speed.
The more depressing observation is that 33.5ECUs are equivalent to 8 cores @ 2.93GHz on Intel's recent architecture. This means your typical "small" EC2 instance with 1 ECU is on a par with ~700MHz of a single modern Intel core. (Highly unscientific but an interesting ballpark.)
Amazon says that "one EC2 Compute Unit (ECU) provides the equivalent CPU capacity of a 1.0-1.2 GHz 2007 Opteron or 2007 Xeon processor"[1].
> > The more depressing observation is that 33.5ECUs are equivalent to 8 cores @ 2.93GHz
You said:
> I don't know where you got 700 MHz from, because by my math, it'd be equivalent to a 1.43 GHz Nehalem core (33.5 / 8 * 2.93)
Transposed.
>>> 8*2.93/33.5
0.69970149253731351
8 cores at 2.93 GHz is 23.44 GHz, which means one compute unit is 700 MHz.
A modern processor can do more per clock than an older processor. In addition, it has a larger cache, faster cache, and a faster memory bus, although on the flip side the memory bus is being shared between more CPUs.
P.s. Disclaimer, they are friends of mine, the beta is pretty epic!
It seems like they will never turn evil, but most big companies do, or perhaps they are just hiding it very well. ;)
Now if only I had some use for this :( (inspiration welcome, I am writing about GPU programming right now).
What is the mapreduce of CUDA going to be?
And the pricing... to quote from the other article[1] on the GPU instances that's on the front page right now:
"An 8 TeraFLOPS HPC cluster of GPU-enabled nodes will now only cost you about $17 per hour."
[1] http://www.allthingsdistributed.com/2010/11/cluster_gpu_inst...
Is anybody else thinking "custom built rainbow tables for any algorithm you like, delivered directly into your S3 bucket in under 1 hour for $25"?
(I wonder how much this means I should up the bcrypt workfactor to keep my password hashes secure from typical website crackers?)
Depending on how well a problem maps to the massively parallel architecture of a GPU, this may not matter.
http://blog.cyclecomputing.com/2010/11/a-couple-more-nails-i...