I have a continuing itch for more hardware, but I was reminded not long ago that cloud services might be viable to a degree now. This would get around space, heat, noise problems I'd have running it myself.

The economics I've yet to go through in detail, and while I think running your own will still work out cheaper in the long term, I'm more looking at cloud based for short term use (like challenges).

For starters, I've looked at Amazon c4.8xlarge instances, which give you 36 vCPUs at around US$1.675/h. The lower tiers scale with the number of vCPUs so for practical purposes this will do. These are run on E5-2666v3 CPUs, which are Haswell generation so should be fast per clock, although the base clock of 2.6 GHz to turbo at 3.3 GHz is less remarkable. Also I'm concerned about the vCPU, which I presume has hyper threading enabled. That is, I'm only getting half the number of physical cores than I think. Is that the case? As far as I can tell those Xeons can be run in 2 socket systems, for a total of 20 cores/40 threads, hence the maximum of 36 offered per instance I assume. Still, assuming 18 physical cores isn't bad, although I wonder if throughput might be better if I went for multiple smaller ones, since even with quad channel ram, I can't but fear it will be severely ram bandwidth limited.

Then there's actually running it... I have no experience in this. If I understand correctly, you can use the supplied EBS to provide persisting storage. So that, with minimal configuration, I could retain state between running it? I don't want to have to set everything up every time I fire it up. I think I'll have to sign up for the free tier to get some experience.

Is there anything other than Amazon I should look at also? I believe some have used other services not strictly within their TOS, and I don't want to go there. I'd like something that will likely be around for the long term.

Bit of a reply problem up there?

Did a little more digging and just found out about the spot pricing, which isn't what I was looking at earlier. Right now for example, the same c4.8xlarge instance is under $0.30/h! When I looked yesterday, it was around $0.50/h, so variable but a fair portion less than above. My question then is... how does it work? Does your instance just die when you get outbid, or do you get time for a graceful shutdown? Can it be automated in some way e.g. run only if price is <$xx. I've still got more reading up to do.

Haven't had time to look further at this. I would comment that you're supposed to get a 2 minute warning before spot instances are terminated if you should get outbid, so those better at linux than me could automate that to shut down boinc gracefully before then.

I did a rough calculation on the costs of the cloud. Very roughly, running c4.8xlarge at $0.30/h for one year equivalent, would cost in the same ball park as buying an overclocked Skylake computer-only system (no fancy GPU, or other user comforts). I haven't factored in electricity running the Skylake system, nor if any taxes are due on the Amazon pricing (since it is in US$), but even with these crude estimates, I will get more work out of buying systems than giving money to Amazon.

I'm probably too generous on the capabilities of the Amazon offering anyway, since it will have much less memory bandwidth per core than an overclocked consumer CPU. It might offer more L3 cache per core, but it is unclear if this might offer a benefit.

Funnily enough, I almost ordered the parts from Amazon, but they didn't have the CPU in stock so I found elsewhere.

So as far as I'm concerned, going cloud is on hold indefinitely as the economics for long term running don't stack up. I haven't ruled it out as something for challenges, but I probably wont be motivated enough to have a go at that until closer to the next specific challenge.

I did a rough calculation on the costs of the cloud. Very roughly, running c4.8xlarge at $0.30/h for one year equivalent, would cost in the same ball park as buying an overclocked Skylake computer-only system (no fancy GPU, or other user comforts). I haven't factored in electricity running the Skylake system, nor if any taxes are due on the Amazon pricing (since it is in US$), but even with these crude estimates, I will get more work out of buying systems than giving money to Amazon.

I'm probably too generous on the capabilities of the Amazon offering anyway, since it will have much less memory bandwidth per core than an overclocked consumer CPU. It might offer more L3 cache per core, but it is unclear if this might offer a benefit.

Funnily enough, I almost ordered the parts from Amazon, but they didn't have the CPU in stock so I found elsewhere.

So as far as I'm concerned, going cloud is on hold indefinitely as the economics for long term running don't stack up. I haven't ruled it out as something for challenges, but I probably wont be motivated enough to have a go at that until closer to the next specific challenge.

Look at here
https://developer.ibm.com/linuxone/registration/
It is IBM linux one server: you have 90 days of absolutely free, you dont even need to have credit card.
It works under Linux ( you have few Images of Linux) and can be setup in 10 minutes.
Currently I use it for sieving ,so my cores can do LLR.

____________
92*10^^1439761-1 NEAR-REPDIGIT PRIME :) :) :)
4 * 650^⁴⁹⁸¹⁰¹-1 CRUS PRIME
314187728^¹³¹⁰⁷²+1 GENERALIZED FERMAT
Proud member of team Aggie The Pew. Go Aggie!

Look at here
https://developer.ibm.com/linuxone/registration/
It is IBM linux one server: you have 90 days of absolutely free, you dont even need to have credit card.
It works under Linux ( you have few Images of Linux) and can be setup in 10 minutes.
Currently I use it for sieving ,so my cores can do LLR.

Look at here
https://developer.ibm.com/linuxone/registration/
It is IBM linux one server: you have 90 days of absolutely free, you dont even need to have credit card.
It works under Linux ( you have few Images of Linux) and can be setup in 10 minutes.
Currently I use it for sieving ,so my cores can do LLR.

How do get boinc running on this? it says it's the 390x architecture only and is being a giant pain :(

Look at here
https://developer.ibm.com/linuxone/registration/
It is IBM linux one server: you have 90 days of absolutely free, you dont even need to have credit card.
It works under Linux ( you have few Images of Linux) and can be setup in 10 minutes.
Currently I use it for sieving ,so my cores can do LLR.

How do get boinc running on this? it says it's the 390x architecture only and is being a giant pain :(

After you deploy image, you must compile boinc to run under 390x architecture.
I cannot say is that can be or cannot be done, but even compiling sr1sieve and sr2sieve need knowledge since it must be changed small part of code.
I would not try to run Boinc on those two cores you get, it is pretty slow ( even for sieving)
So even you compile Boinc, you must compile and rest , since there is no LLR for 390x architecture ( as I know)

Look at here
https://developer.ibm.com/linuxone/registration/
It is IBM linux one server: you have 90 days of absolutely free, you dont even need to have credit card.
It works under Linux ( you have few Images of Linux) and can be setup in 10 minutes.
Currently I use it for sieving ,so my cores can do LLR.

I had a quick look at the TOS, and basically they want you to develop on their system. So arguably this could fit in that (you need to port and test it). Still looks like too much effort for the gain than I have time for.