Aws / Google spot compute instances

[RobertFontaine]

Does anyone here have some links or is using the spot market for compute?

It's been a few years since I looked at this and have been wondering whether this a practical way of scaling up for jobs that can be stopped and started.

Is this something that a couple of hundred dollar could buy some serious compute power for a few hours or is this set up for Enterprise buyers?

Are there compute instances that actually make sense for scientific, hpc, or rendering type compute? The last time I looked some of their configurations were essentially toys and it looked cheap until you considered how much compute you were actually buying. ( ie. for what you were buying in gpgpu you were better of buying your own video card.)... Made sense as a test/development environment but not really as compute....

Thanks Robert.

 

[Patrick]

I use spot instances often when I do AWS testing. The big issue is that if your minimum bid is too low, you can have the instance just shutdown on you. There are ways to work around that. You can set your bid price higher but then you might end up paying more, sometimes a LOT more. I have even seen 45-65 min Linux-Bench runs die in the middle of a run. The other thing I am not crazy about is that provisioning is still a bit slower than me just clicking 2 buttons and having a VM running within 15 sec in the cluster.

 

[cperalt1]

You might want to check out what Joyent has to offer for your compute needs. I've only had a dev account with them and bringing up instances is pretty quick and can be done via API. For your needs depending on the workload their Manta product might make sense with Manta being an S3 type of storage but if you need to perform compute you just submit a Manta job and the compute takes place where your objects live so no need to transfer over the network and you are only charged for the time it takes the job to complete. Hope this helps. Big Data Analytics - Object Storage - Joyent