The Scottish Boinc Team

noetus wrote: ↑Thu Jul 06, 2017 5:14 pm Does it really make sense to purchase 6 year old server technology? I imagine the cost savings over more recent hardware gets erased within a year or two if running them constantly. Isn't it better to pay this extra cost for more recent hardware that will be more powerful as well as more energy efficient?

You have to carefully select your Xeons in each generation. TDP varies wildly, together with the MHz.
Socket 2011 (DDR3) doesn't suffer much compared to the newer Socket 2011-3 (DDR4).

From my testing running Carlos' app, 14 threads on a Xeon V4 clocking at 2.8Ghz with 28GB RAM reserved runs about the same as a 2x X5675 (24 threads) using 36GB RAM. A 10c/20t V4 ES system clocking at 2.4Ghz on all threads should do as much. I'd estimate $1,200 to build a single cpu system w/32GB RAM, maybe $100 more for 12c/24t system. That's if you buy it here in the US though. If the parts can be bought and shipped to the UK for £1,000, then it would be worth it. Would use 50%-66% of the lecky a Z600 2x X5675 system uses (~350 watts?). If can save 150 watts, then assuming your power rate is .15 per kwh, you could save around 197 per year.

Mains electricity costs vary in the EU. In US dollars (using Eurostat figures for 2016), in the UK it is 0.20 per kWh, 33% more than what you were basing your calculations on, which is significant. So the payoff for using more recent hardware is considerably more. In Denmark and Germany the price is even higher - 0.35 and 0.34 respectively, more than double what you were assuming! Just using your figures over 2.5 years we are talking about a cost saving of almost the original cost of the system. No wonder the big tech companies turf out all their old systems after a few years and invest in the latest hardware.

Hungary (where I live) and Bulgaria are the cheapest EU countries for power, 0.13 and 0.11 US dollars respectively. So the payoff for newer hardware partly depends where you live. You also have to factor in the higher chance of hardware failure for older equipment, especially server equipment that has seen 24/7 use (after 6 years some equipment, such as SSDs, might be reaching the end of their design life).

The cheapest place in Europe is Kosovo, a mere .07.

A quick glance at electricity prices in the US (from the US EIA) reveals wide state-by-state differences there too, from .27 in Hawaii to .09 in Washington, with prices generally lower than in the EU.

Another factor to consider is that if you buy one of those old servers you're going to get yourself a pretty noisy system, in all likelihood. If you build a system yourself you can make it near silent. And don't forget - you'll have manufacturer's warranties on all those new components of your build (excluding the CPU if you get a used/ES/QS version). I don't think you get any warranty purchasing used server equipment on eBay, maybe DOA that's all.

What is Carlos' app? I wouldn't mind running some tests of my own.

Guys,

I'm an energy consultant and I trying to understand what's the most efficient machine to run several things like prime hunting, factoring (msieve, nfs@home) and prime gaps.

For example, the new intel desktop cpu's throttle when you use avx512.
So prime hunting uses avx, avx2.0 and in the near future avx512. When using Prime95/LLR with lots of cores the memory is the bottleneck.

Factoring (msieve client, or NFS@Home) is pure GHz needed as well as the gaps prime application.

What I don't understand is the example of my ivybridge laptop when running prime gaps on 4 threads (30 W, 18-20e9 n/s, depends on the range being processed) is more efficient than these systems:

6e9 n/s i3-530 2.9GHz 3 thread (Windows/Cygwin)
19e9 n/s E5-2666v3 2.9GHz 8 thread
27e9 n/s i7-4770k 4.3GHz 8 thread
32e9 n/s i7-6700k 4.2GHz 8 thread
34e9 n/s E5-2666v3 2.9GHz 16 thread
59e9 n/s E5-2686v4 2.7GHz 48 thread

My laptop, sandybridge i7- 3630QM.
Carlos

Prime gaps application:



gap5_g -n1 9000e15 -n2 9001e15 -gap 1346 -delta 155 -sb 24 -bs 18 -t 4 -mem 12

Windows:

start /low /min gap5_g.exe -n1 9000e15 -n2 9001e15 -gap 1364 -delta 155 -sb 24 -bs 18 -mem 12 -t 4

-t number of threads
-mem memory allocated

I'm also considering the prime gaps is not well optimised for full threads or it has memory bottleneck. I suppose it will be quicker on a multi core machine run several instances of the client even set to a lower 1 thread.

pinhodecarlos wrote: ↑Fri Jul 07, 2017 12:27 pmWhat I don't understand is the example of my ivybridge laptop when running prime gaps on 4 threads (30 W, 18-20e9 n/s, depends on the range being processed) is more efficient than these systems:

6e9 n/s i3-530 2.9GHz 3 thread (Windows/Cygwin)
19e9 n/s E5-2666v3 2.9GHz 8 thread
27e9 n/s i7-4770k 4.3GHz 8 thread
32e9 n/s i7-6700k 4.2GHz 8 thread
34e9 n/s E5-2666v3 2.9GHz 16 thread
59e9 n/s E5-2686v4 2.7GHz 48 thread

Make sure each thread has the same amount of RAM, when you make a comparison like this.
IMHO a 48 thread Xeon should have 192GB of RAM (4GB/thread), a 16 thread Xeon 64GB, a 8 thread Xeon or i7 32GB.

Verify how the different options affect performance...

Usage: gap [options]

Options:
-n1 x First number to check is x
-n2 y last number to check is y
-gap g searching for gap>=g (default=1346)
-delta d sieving on k*m+[0,d) intervals (default=196)
-sb u sieve uses 2^u bits of memory (default=25)
-bs v one bucket size is 2^v bytes (default=14)
-mem m the maximal memory usage is m GB (m can be any real number)
-t k use k threads

Maybe different values for -gap, -delta, -sb, and -bs give better performance on higher thread or memory systems.

We are all using the same gap, delta, sb and bs sometimes.

Reviving and old thread. Soon my wife will start her PhD which will be based on model simulation. I hope she will have access to high end computers to run gprooms or aspen, otherwise we will have to purchase a machine. I suppose both softwares will run better with higher number of threads and lots of memory. Could anyone advise me on tower racks, budget £1200. Anyone here is an U.K. is seller?

This is much appreciated.

Carlos

Tried www.bargainhardware.co.uk ?