Renting huge single CPU/memory servers?

[ranakor]

Hello there,
My google-fu turned me down here as it's very hard to google for renting servers short term (like you'd do for a rendering job on a farm), not for rendering (i want to run my own software, get a remote desktop as admin etc), on very very powerfull servers (the better the merryer) that are actually single instance windows servers (not farms / clusters / whatever).
Does anyone know of a company renting short term (days, or a month at most) very large (talking 80 threads minimum with 256GB ram minimum) servers and if so at what price? At first i thought i wouldn't find a company doing this because the prices for those servers must be prohibitive but checking around it seems possible to make such a server for a reasonable cost so i'm really hoping i can save up and find a company renting that instead of getting one made.
Thanks!

 

[dma0991]

Amazon EC2

 

[ranakor]

Amazon EC2 is exactly NOT what i'm looking for as i explained, i have no use for even 2 bajillon gigaherz if they're spread around in machines / instances / cloud, i want a single physical machine with a single physical motherboard to launch a single .exe on a non networked setup.
If looking at a single instance even the largest EC2 instances are very very low end.

 

[MisterMac]

What your looking at is 4P dense servers, which are ouchies expensive.

You could probably get short term yearly contracts, but that's what your looking at - short of some local hosting firm helping you out.


Softlayer and other big "quantity" hosters will probably have some solutions for you.
Altho if you can find some offering opteron servers - and your app truly is multihreaded that might be heavily worth it.


EDIT:
Don't know your story or resources - but if your in EU i'd highly recommend some online server e-tailers and build it yourself.
It's gonna be far cheaper in the length.
short of 1000e a month for a 4 proc dense server.

 

[MisterMac]

I see softlayer offers some hexcore\octocore - but else your looking at custom orders from a company.

It will be cheaper if your want 80 threads and 256 GB of ram - at minimum to build yourself.

The mainstream "server" market does not need this much density by far.
Get your program to be clustered and you'll be alot more happy

 

[Idontcare]

Are you looking for something like a VPS?

 

[MisterMac]

Are you looking for something like a VPS?

Since he declined EC2 - that'd probably be a big no


I'm quite wondering what his workload is, maybe OP Can post what his workload is.

 

[Idontcare]

Since he declined EC2 - that'd probably be a big no


I'm quite wondering what his workload is, maybe OP Can post what his workload is.

His request in the OP reads exactly like what we use VPS for in the forex business, but the VPS we rent do not require you to install and run forex software, you can run anything you want to install.

Hello there,
My google-fu turned me down here as it's very hard to google for renting servers short term (like you'd do for a rendering job on a farm), not for rendering (i want to run my own software, get a remote desktop as admin etc), on very very powerfull servers (the better the merryer) that are actually single instance windows servers (not farms / clusters / whatever).

^ that is precisely what you get from a VPS that is being targeted towards the forex demographic.

 

[ranakor]

The hightest i've found so far is 512GB ram / 4X10cores @ softlayer which would run at around 10K / month, considering i'd only need this once in a while and they don't seem to have a setup fee that would be a good deal.
Anyone can think of something more powerfull without having to purchase?

@MisterMac : i don't own the program but it wouldn't be possible to get it clustered anyway and preserve the performance. The calculation are parallelizable but the data used or resulting from those calculations isn't (each thread for each step requires the whole data in memory). The workload is generating very large and detailed worlds from complex graphs in world machine WITHOUT using the tiled export (which doesn't give the same results). While at low resolution it generates in seconds it scales up very very quickly as you up the resolution and i'm thinking what i'm looking at would already take quite a while with a very large server.

@Idontcare : no not a vps, a vps is typically much slower than even a low end dedicated server.

Edit @ MisterMac : yes it would be cheaper to purchase than to rent for a year, but i'm not planning to need it for a year. However i'm still in the thinking phase for all this so decision really depends on the quotes i'll get from my research. 8 way 10 core + HT supermicro server is sounding very tempting, but i'm guessing it's pretty pricey

 

[Ferzerp]

The hightest i've found so far is 512GB ram / 4X10cores @ softlayer which would run at around 10K / month, considering i'd only need this once in a while and they don't seem to have a setup fee that would be a good deal. Anyone can think of something more powerfull without having to purchase?

If you have the power to run this yourself (1.5ish kW circuit size needed probably), the same server will run you only around $30k-$35k. If you're buying in 1 month increments, and need it 3 times, you can just buy one yourself as well as being able to depreciate it as capital instead of it being an operating expense (assuming it is business related).

 

[MisterMac]

His request in the OP reads exactly like what we use VPS for in the forex business, but the VPS we rent do not require you to install and run forex software, you can run anything you want to install.



^ that is precisely what you get from a VPS that is being targeted towards the forex demographic.

a VPS is virtualized.


EC2 is virtualized.

Anything virtualized has physical limits - and OP was describing the OUTER limit of physical box requirements, hence VPS makes no sense for anyone to offer him market wise.

Short of a bladecluster with shared memory busses - he's looking at dense 4unit 4proc heatboxes.
Which are expensive for that specific reason - the power in physical size is unmatched.

ranakor:

What i would do is - test how well threaded your program is - 24 hour run somehow.

Writing software that can be "n Threaded" many "do" - but making it work effectively is nearly impossible without some hardcore optimization of locks.
See where the borderline is - and get a xeon with the highest clocks possible at your threadlevel.

there might be some interesting experiments in getting your own chassis as long as you don't mind building\installing yourself.

If your workload comes from academic nature - there's bound to be very nice discount @ hp\supermicro for some chassis with quad processor capable boards.

If thread and not clock is king - even interlagos 16 core options would maybe be interesting.


In short, your can save a ton of money from buying rather than renting.
Even if you go to those professional enterprise grade hosters - they're fees alone will probably set you back as much as Softlayer or another "big-box" hoster.

 

[MisterMac]

Quick another post - if you set some goals cost is easier to calculate.

But some Fully equipped 2u\quad proc chassis with 8\10 core 7500\4800 Xeons would probably cost around 30.000 USD roughly with controllers and SAS disks.
(assuming my calculations arent completely wrong).

Atleast in the US + whatever you can get of academic discount.

 

[ranakor]

The work isn't academic in nature, it's professional work in my company.
Also multithreading complexity is often missunderstood, it's not complex if you're accessing immutable data and as far as i can guess, world machine is definately not mutating the inputs when computing an output, so it shouldn't require any locking if properly done. I don't know how well the data layouts and algorithms are optimized for cache misses & co but in any case world machine is definately many threads compatible and has already run for me on 16+ threads over way more than 24 hours with no issue.
The program do is very well parallelized with rather little diminishing returns, i can't tell about going as far as 160 threads but the author has success stories from people running it over 16+ cores efficiently and the nature of the graphs i'll be making should be particularily well parallelizable. Same for the memory, memory usage in world machine increases rather quickly with resolution so while i still need to figure out the decent core/memory ratio i don't think 512GB would end up being too much.
For the 30K-35K quotes, do you have a list of parts? It seemed to me that it would be a lot more expensive to get something similar to what i mentioned at softlayer (4X10 cores + 512GB ram). Disk price is pretty meaningless, i don't need massive storage at all, just 2 decent size SSD in raid 0 will do i'd think
Power is not an issue, i can just put it in a colocation if i purchase, i'm not going to be running it locally.

 

[MisterMac]

Call dell & hp - get an offer.

Call them back 2 days after - tell them hp\dell offered it much cheaper.
See how low you can get.


A base chassis HP with 2U 500 series, some ram, no disks and a nehalemEX should be around... 15k to 25k (After your little rebate hustle discussion - DELL is more prone to cheaping out here FYI, i'm just a HP fanboi in the server world).

That includes some base ram like 128 GB, NIC and a simple raidcontroller - with one of the cheaper 7500 low clocked models ive bought in the past for 19.000 even including free LF SAS disks(Before thailand - the stock pool was crazy and everyone wanted to move units).

All you'd need to decide price wise is:

- High clocked or low clocked?
- 8/16 or 10/20 ?
- Lots of L3 cache or not?

If you say yes to all - you'll quickly quickly head up to 30k-40k pr. 2u rack.
Which is why i'd go bench your app in small mainstream tests where you figure out - where's the rough scaling points.

Raw threads? does cache matter? clockspeed over threads? etc.

You can have a cheap 15.000 outfit, if you going 8/16 with low clocks and low cache x 4.
Add 5000 each time you take a step up with clocks and cache.

Nehalem EX based boards\chassis (6500/7500) for 4p servers could probably be had cheaper than westmere ex ( 4800/8800) tho.


I would start calling even if your intent to buy is far ahead.
Complete newbiz for most 4p vendors\resellers - is treated like royalty since it rarely occurs outside each local markets players.

 

[MisterMac]

Looking at WorldMachine - i'm not really overly confident you will find a user for 80 threads - perhaps lots of ram if you as you said run into scale issues that's just pure dataspace.


From reading it - i'd suggest it's tiered in it's "Mulithread" approach - where several branches to intensive workloads is a single threaded path.

Also it references optimally 16 threads several places and in their forum.


You need to test this on some 12/16 thread machine with the same clocks\cache\uarch - to see the advantage first imho.

 

[ranakor]

Am i not going to get much highter prices from dell/hp than i would from something like supermicro?
Keep in mind i just need to process very long but not mission critical jobs, so while the server would run at full power and must be stable enought for this i don't want to pay a huge premium for support / warranty / 8hour replacement part etc, if it fails then i really don't care if it's down for 8 hours or 3 days for exemple.
Edit: it's already been tested on 16 threads and gives linear improvement per thread so far, so scaling should be close to ghz*threadcount at equal processor architechture