mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > Hardware > Cloud Computing

Reply
 
Thread Tools
Old 2019-03-28, 08:58   #1
M344587487
 
M344587487's Avatar
 
"Composite as Heck"
Oct 2017

22×163 Posts
Default PRP testing service

I'm toying with the idea of building a small GPU farm for PRP testing. I want to fondle some hardware but there's no way I can afford the electricity cost or initial outlay. Is there any interest in a PRP testing service? First time PRP only as not every PRP type can be DC'd by gpuowl. I'm thinking of something simple:
  • Pay per test
  • Send me worktodo lines
  • I send back results.txt and interim files. You add the results manually to primenet from your account
  • You get credit if a prime is found
  • If a test is proven to have errors you get another similarly-sized test at no cost
Pricing would scale with the square of the exponent mod 1000000 for simplicity. I think the idea is viable at $5 for an 85M test, which would scale like this:
Code:
80M $4.43
81M $4.54
82M $4.65
83M $4.77
84M $4.88
85M $5.00
86M $5.12
87M $5.24
88M $5.36
89M $5.48
90M $5.61
91M $5.73
92M $5.86
93M $5.99
94M $6.11
95M $6.25
96M $6.38
97M $6.51
98M $6.65
99M $6.78
.
.
.
332M $76.28
333M $76.74
With enough demand this would pay for the hardware and running costs while providing a cheaper alternative to the cloud. Food for thought.
M344587487 is online now   Reply With Quote
Old 2019-03-28, 09:04   #2
SELROC
 

68016 Posts
Default

Quote:
Originally Posted by M344587487 View Post
I'm toying with the idea of building a small GPU farm for PRP testing. I want to fondle some hardware but there's no way I can afford the electricity cost or initial outlay. Is there any interest in a PRP testing service? First time PRP only as not every PRP type can be DC'd by gpuowl. I'm thinking of something simple:
  • Pay per test
  • Send me worktodo lines
  • I send back results.txt and interim files. You add the results manually to primenet from your account
  • You get credit if a prime is found
  • If a test is proven to have errors you get another similarly-sized test at no cost
Pricing would scale with the square of the exponent mod 1000000 for simplicity. I think the idea is viable at $5 for an 85M test, which would scale like this:
Code:
80M $4.43
81M $4.54
82M $4.65
83M $4.77
84M $4.88
85M $5.00
86M $5.12
87M $5.24
88M $5.36
89M $5.48
90M $5.61
91M $5.73
92M $5.86
93M $5.99
94M $6.11
95M $6.25
96M $6.38
97M $6.51
98M $6.65
99M $6.78
.
.
.
332M $76.28
333M $76.74
With enough demand this would pay for the hardware and running costs while providing a cheaper alternative to the cloud. Food for thought.



Are you saying someone will pay you instead of paying electricity bill ?


I can do it, I have enough GPUs to do it.

Last fiddled with by SELROC on 2019-03-28 at 09:05
  Reply With Quote
Old 2019-03-29, 06:28   #3
M344587487
 
M344587487's Avatar
 
"Composite as Heck"
Oct 2017

22·163 Posts
Default

It's an idea people have toyed with for years but it's not something that's materialised at least not publicly. Someone should do it at least to undercut the cloud which seems ill-suited to PRP yet is currently being used to some degree. I'm willing to to make a proper go of it and ramp up to whatever the demand is.
M344587487 is online now   Reply With Quote
Old 2019-03-29, 06:37   #4
SELROC
 

3·449 Posts
Default

Quote:
Originally Posted by M344587487 View Post
It's an idea people have toyed with for years but it's not something that's materialised at least not publicly. Someone should do it at least to undercut the cloud which seems ill-suited to PRP yet is currently being used to some degree. I'm willing to to make a proper go of it and ramp up to whatever the demand is.

The idea is not bad. You could set up a batch scheduler, where you admit your customers to submit jobs to your computers.

Last fiddled with by SELROC on 2019-03-29 at 06:38
  Reply With Quote
Old 2019-03-29, 11:01   #5
a1call
 
a1call's Avatar
 
"Rashid Naimi"
Oct 2015
Remote to Here/There

36038 Posts
Default

IMHO There is a fortune to be made by small scale data centres which are willing to lease out their computers or the computers are regular people who are willing to use them as a third party to monetize their idle clock cycles.
The trick is to rate the lease equivalent to a car rental ratio say $50 a day for 40k worth of computers.
Or much less, since computer depreciation is much less than cars and you don't need to use Brand new computers.
As is you will end up paying for the equivalent computer in couple of weeks for services offered by the big 3 Cloud computing providers which makes no sense to me.

Electricity is really not necessarily a factor in many locales.
I pay about $800 a month to heat my place during winter months. Any computing generated heat/cost is a non issue in Canada.

Last fiddled with by a1call on 2019-03-29 at 11:35
a1call is offline   Reply With Quote
Old 2019-03-29, 14:51   #6
SELROC
 

11001011100012 Posts
Default

Quote:
Originally Posted by a1call View Post
IMHO There is a fortune to be made by small scale data centres which are willing to lease out their computers or the computers are regular people who are willing to use them as a third party to monetize their idle clock cycles.
The trick is to rate the lease equivalent to a car rental ratio say $50 a day for 40k worth of computers.
Or much less, since computer depreciation is much less than cars and you don't need to use Brand new computers.
As is you will end up paying for the equivalent computer in couple of weeks for services offered by the big 3 Cloud computing providers which makes no sense to me.

Electricity is really not necessarily a factor in many locales.
I pay about $800 a month to heat my place during winter months. Any computing generated heat/cost is a non issue in Canada.

fresh news:


https://www.phoronix.com/scan.php?pa...acket-Roll-Out


https://amperecomputing.com/ampere-a...dge-workloads/

Last fiddled with by SELROC on 2019-03-29 at 14:52
  Reply With Quote
Old 2019-03-29, 15:17   #7
ATH
Einyen
 
ATH's Avatar
 
Dec 2003
Denmark

3·23·43 Posts
Default

32 cores for $1/h is not very cheap, right now the 36 cores on EC2 with 144 GB RAM and 2x900 GB SSD is $0.6841 per hour.
ATH is offline   Reply With Quote
Old 2019-03-29, 16:05   #8
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

10001111001012 Posts
Default

Quote:
Originally Posted by M344587487 View Post
I'm toying with the idea of building a small GPU farm for PRP testing. I want to fondle some hardware but there's no way I can afford the electricity cost or initial outlay. Is there any interest in a PRP testing service? First time PRP only as not every PRP type can be DC'd by gpuowl. I'm thinking of something simple:
  • Pay per test
  • Send me worktodo lines
  • I send back results.txt and interim files. You add the results manually to primenet from your account
  • You get credit if a prime is found
  • If a test is proven to have errors you get another similarly-sized test at no cost
Pricing would scale with the square of the exponent mod 1000000 for simplicity. I think the idea is viable at $5 for an 85M test, which would scale like this:
Code:
80M $4.43
81M $4.54
82M $4.65
83M $4.77
84M $4.88
85M $5.00
86M $5.12
87M $5.24
88M $5.36
89M $5.48
90M $5.61
91M $5.73
92M $5.86
93M $5.99
94M $6.11
95M $6.25
96M $6.38
97M $6.51
98M $6.65
99M $6.78
.
.
.
332M $76.28
333M $76.74
With enough demand this would pay for the hardware and running costs while providing a cheaper alternative to the cloud. Food for thought.
Interesting. At the top end of the mersenne.org range, 999M, $690.66/each.
My tests (both gpu and prime95) show a p2.095relationship, which I often round to 2.1 power. So the schedule above seems to offer a bit of a discount for larger p.

$5/84M is a bit below current cloud costs recently posted.
It is however, more than double what my computed best gpu cost is ($2.12/84M raw, $1.78 with heating benefit, gpuowl on RX480), about double what my somewhat aged best cpu performers cost, and also higher than all my GTX10xx models running CUDALucas 2.06. Among my gpus, only my ancient GTX480 and Quadros cost more than your proposed rate. Financial breakeven heavily depends on electrical rates and hardware cost amortization rate. You would have a strong financial incentive to tune for most cost-effective throughput.

If you see a prime result before the paying client does, usual rules, first human to know it, would name you as the discoverer. Perhaps you plan to employ some fancy scripting so the client sees the results and you do not?

What payment methods do you anticipate using? It appears to be payment in advance, judging by your list order.
Is it a first-in-first-out service? (One could consider offering a higher cost "express lane")

How do you plan to keep track of multiple clients' runs on multiple gpus?

84M on my RX480 takes ~85 hours, so about 2/week. $5 charge vs $2.12 cost x 168/85/week x 52wk/yr x 5 gpus would be ~$1480. net profit annually (more than the cost of the new gpus). So if it's a labor of love, within power dissipation limits, it could pay to grow itself briskly. Unless someone else undercuts your pricing.

If considering using a UPS, be careful of its efficiency. I've seen efficiency figures approaching 99% and as low as 84%.

The error free guarantee is a nice touch. It could take years to have effect though. (LL DC is lagging LL by about 10 years. https://www.mersenne.org/report_expo...6753351&full=1) And that means the rerun is done with much faster gpus, at less cost.

If you're planning to also run your own PRP tests, or perhaps regardless, I recommend segregating them to a separate system, and keeping good logs permanently, for possibly defending any legal disputes that might arise. (I once was asked by my supervisor for data regarding a complaint lodged to someone several levels further up in the state of WI org chart, from an unhappy vendor complaining about not having been awarded a contract or offered a chance to bid. It turned out that they had been included in the bidder list, asked to bid, and replied "no bid"! Good records can be very valuable.)

Last fiddled with by kriesel on 2019-03-29 at 16:26
kriesel is online now   Reply With Quote
Old 2019-03-29, 16:25   #9
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

32×509 Posts
Default

Quote:
Originally Posted by M344587487 View Post
Pricing would scale with the square of the exponent mod 1000000 for simplicity.
I'm guessing you meant something like floor(exponent/1000000) there.
So that any one 80000000<p<81000000 is one price, listed as 80M.
Pricing function should be a staircase for simplicity. Your example prices seem consistent with floor and staircase.

Not a sawtooth function, generated by mod, where p just above a multiple of 106 is nearly free, and just under a multiple of 106 is very costly compared to a slightly higher p that's just over a multiple of 106.
kriesel is online now   Reply With Quote
Old 2019-03-29, 19:31   #10
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

32×509 Posts
Default

Quote:
Originally Posted by ATH View Post
32 cores for $1/h is not very cheap, right now the 36 cores on EC2 with 144 GB RAM and 2x900 GB SSD is $0.6841 per hour.
That's just one of many offerings, some ARM, some Intel, listed at https://www.packet.com/cloud/servers/ Do any look competitive? It would take some doing to translate these from $/hr to $/84M primality test.
Poking around a bit, I see spot prices there in some cases much lower than the hourly rate (25%, 20%, lower). Reserved rates offer discounts 5-40%. Wide variety of linux flavors and here and there Win2016 standard or 2012. 28 Core Epyc $.25/hr spot;

https://www.packet.com/cloud/servers/m1-xlarge/ 24 core Intel e5-2650 $1.70 hourly, 0.25/hr spot.
Still looks steep though compared to my dual-e5-2690 at $.05/hour.

Last fiddled with by kriesel on 2019-03-29 at 20:02
kriesel is online now   Reply With Quote
Old 2019-03-29, 19:57   #11
M344587487
 
M344587487's Avatar
 
"Composite as Heck"
Oct 2017

22×163 Posts
Default

Quote:
Originally Posted by kriesel View Post
Interesting. At the top end, 999M, $690.66/each.
My tests (both gpu and prime95) show a p2.095relationship, which I often round to 2.1 power. So the schedule above seems to offer a bit of a discount for larger p.
The limited testing I've done for this card puts the scaling closer to 2 but I wasn't exhaustive so the figure could be off. Encouraging bigger tests makes it easier to fill capacity but really I just went for simple. You could also min-max by requesting exponents near the end of an xxM block.

Quote:
Originally Posted by kriesel View Post
$5/84M is a bit below current cloud costs recently posted.
It is however, more than double what my computed best gpu cost is ($2.12/84M raw, $1.78 with heating benefit, gpuowl on RX480), about double what my somewhat aged best cpu performers cost, and also higher than all my GTX10xx models running CUDALucas 2.06. Among my gpus, only my ancient GTX480 and Quadros cost more than your proposed rate. Financial breakeven heavily depends on electrical rates and hardware cost amortization rate. You would have a strong financial incentive to tune for most cost-effective throughput.
$5 per 85M test is roughly breakeven after one year at 100% capacity assuming no failures, minimal downtime and tuned for the most cost-effective throughput for a small farm of up to ~$4800 outlay. 100% capacity is unlikely and there's a degree of risk and time which is factored in. Once breakeven has been met and it hopefully proves to be relatively painless to administer I'm not against dropping the price to closer to cost but I also don't value the effort of maintaining the service at zero. Prime hunting as a hobby my time is freely given but tying myself to a commitment and being accountable to others makes it another matter. Lowest cost per test will always be owning your own hardware unless the cloud gets a whole lot cheaper.

I had thought about proposing the idea as a community-driven effort but that puts too much trust and responsibility on me and too many cooks can make things difficult. This way it can be scaled to whatever demand and my ability dictates and basically lets the market decide if it lives or dies. Community efforts sometimes lead to resentments and high school drama especially where money is concerned, seems wise to sidestep that.

Quote:
Originally Posted by kriesel View Post
If you see a prime result before the paying client does, usual rules, first human to know it, would name you as the discoverer. Perhaps you plan to employ some fancy scripting so the client sees the results and you do not?
That's a funny rule, what happens if the result is automatically sent from an unattended PC and a database mod sees the result before the owner of the hardware? I'd already planned on sending the results to the user and having them report to primenet, it wouldn't take much to have the results sent to the user by email with no interaction on my part.

Quote:
Originally Posted by kriesel View Post
What payment methods do you anticipate using? It appears to be payment in advance, judging by your list order.
Is it a first-in-first-out service? (One could consider offering a higher cost "express lane")
Payment first, first-in-first-out. I doubt there's demand to jump the queue, the only time that makes sense is a quick DC for a prime candidate and honestly I'd do that anyway if it was brought to my attention. Paypal I guess. Cryptocurrency of any popular flavour. UK to UK bank transfer. I'm assuming international bank payments would be a nightmare.

Quote:
Originally Posted by kriesel View Post
How do you plan to keep track of multiple clients' runs on multiple gpus?
Depends on the scale. Assuming the scale is small with a few users ordering many exponents it could by done by spreadsheet. It wouldn't be much effort compared to the real ballache which is managing payments. A slightly larger scale could be managed with a script which I'd babysit along with the hardware, the simplest option at a glance is a directory per exponent to avoid mixing results.

Quote:
Originally Posted by kriesel View Post
84M on my RX480 takes ~85 hours, so about 2/week. $5 charge vs $2.12 cost x 168/85/week x 52wk/yr x 5 gpus would be ~$1480. net profit annually (more than the cost of the new gpus). So if it's a labor of love, within power dissipation limits, it could pay to grow itself briskly. Unless someone else undercuts your pricing.
Anyone is free to undercut and they could easily do so particularly if they already own the hardware. The price I've set factors in the anticipated risk and time commitment I'm willing to bear. That said I'm principally suggesting this to acquire hardware to play with not profit as such, see above.

Quote:
Originally Posted by kriesel View Post
If considering using a UPS, be careful of its efficiency. I've seen efficiency figures approaching 99% and as low as 84%.
Thanks. I am considering a UPS and didn't know efficiency was a major factor. A UPS isn't a strict requirement as the supply is pretty stable and the workload not time critical, but it would be nice.

Quote:
Originally Posted by kriesel View Post
The error free guarantee is a nice touch. It could take years to have effect though. (LL DC is lagging LL by about 10 years. https://www.mersenne.org/report_expo...6753351&full=1) And that means the rerun is done with much faster gpus, at less cost.
Perhaps, the guarantee is to try to be fair I didn't consider the landscape a decade from now. The Radeon VII is such a beast at PRP and unlikely to be made in huge numbers that the only thing I can see that may beat it in a cost-effective way within 3 years say is Navi or Radeon VIII if it'll even exist. I could be wrong but think that Navi is next gen Polaris and may beat Vega 20 in enough numbers, but there is added packaging, logistics and power cost in using many smaller GPUs that don't make it a certainty. Unless some clever sod utilises int and fp simultaneously on the newer nvidia cards, or nvidia seriously rethinks their DP policy, I don't see nvidia cards being price-competitive at PRP anytime soon (don't follow nvidia closely enough to know for sure). 10 years is a long time though.

Quote:
Originally Posted by kriesel View Post
I'm guessing you meant something like floor(exponent/1000000) there.
So that any one 80000000<p<81000000 is one price, listed as 80M.
Pricing function should be a staircase for simplicity. Your example prices seem consistent with floor and staircase.
Yep that's what I meant, integer division not the modulus my mistake.
M344587487 is online now   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Factoring as a service Nick Factoring 8 2016-03-18 03:37
LLRNet as a NT service rover Sierpinski/Riesel Base 5 6 2006-10-02 13:43
Speed of P-1 testing vs. Trial Factoring testing eepiccolo Math 6 2006-03-28 20:53
How to fix the Prime95 Service E_tron Software 1 2003-10-05 03:55
v22.9 service problem PageFault Software 2 2002-09-26 01:35

All times are UTC. The time now is 09:45.

Thu Oct 22 09:45:44 UTC 2020 up 42 days, 6:56, 0 users, load averages: 1.76, 1.63, 1.68

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2020, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.