Bargain Hardware Thread

[Madpoo]

Originally Posted by VBCurtis

Thanks, Aaron! If M1277 falls to my ECM, you deserve credit for making it possible.

I'm glad that's one of the things you'll be using it for. I (temporarily?) gave up my own ECM hunt on M1277 since I was focusing resources on the strategic double-checking, but I still think that was a fun side-project.

[Gordon]

Originally Posted by VBCurtis

Madpoo turned out to have *much* more than dozens of 1GB sticks. He mailed me a complete set of 32 sticks of 4GB DDR2 ECC, filling the server with 4 nodes at 32GB each. The memory is now installed, and will be used to run large-bound ECM curves on M1277 on a thread or two while the rest of the nodes run NFS tasks.

Thanks, Aaron! If M1277 falls to my ECM, you deserve credit for making it possible.

32gig doesn't go far, with B1=800M

B2=3e14 needs 17 gig

B2=4e14 needs 34 gig

Think you might be disappointed...

[VBCurtis]

Those B2's are default matched up with B1 = 4e9 and up, which is quite large. I have no idea why you are quoting such B2 values with 800M!

I have chosen to run B1 = 4.5e9 on P95 and B2 = 150e12 (k = 3, so I could go higher but expected t70 factor time rises if I do), which indeed requires 17GB for Stage 2. One thread of stage 2 per node means roughly 25 curves per day. These settings are within 2% of the best expected time for t70 (I tested 100M increments of B1 from 9e8 to 6e9). 15000 curves for a t65, 69000 for a t70.

I learned the power supplies are not strong enough for these servers- I had to turn memory down to 533 from 667 (using a "low power" setting in BIOS) and remove 8GB from two of the nodes to prevent the power supplies from shutting down when all 32 cores are in use. I also removed all-but-one hard drive from each node; my kill-a-watt suggests the 500w-rated power supplies shut down when wall plug draw exceeds 460w. Too bad they're rather non-standard supplies..

[xilman]

Originally Posted by Gordon

32gig doesn't go far, with B1=800M

B2=3e14 needs 17 gig

B2=4e14 needs 34 gig

Think you might be disappointed...

People, please!

Learn how to use GMP-ECM with -maxmem

It's how I run seriously high B2 on memory constrained processors. Which would you rather have: a 10% performance drop by constraining memory usage or a 100% performance drop by not doing so?

Example:

pcl23@brnikat:~/ls/nums$ ps ax | grep ecm
13707 pts/3    R      0:15 ecm -maxmem 1024 850000000
13711 pts/3    R      0:08 ecm -maxmem 1024 850000000
13713 pts/3    R      0:03 ecm -maxmem 1024 850000000
13717 pts/3    R      0:01 ecm -maxmem 1024 850000000
13719 pts/3    R+     0:00 grep ecm
pcl23@brnikat:~/ls/nums$ head -1 /proc/meminfo
MemTotal:       16311512 kB
pcl23@brnikat:~/ls/nums$

Four ecm processes with B1=850M running on a 16G machine with plenty left over for Real Work™

[lavalamp]

I have found that GMP-ECM seems to somewhat overestimate the amount of memory it will use, so I have shied away from using -maxmem. Instead I have preferred to use -k and -treefile to manually tune it to the system I am running on.

To give a particular example, on the number 10^999 + 13, I ran GMP-ECM with B1=2.9E9 and B2=1E14, and with the options -k 5 and -treefile. GMP-ECM quoted an estimated 20GB memory usage, yet never used more than about 15GB.

In reply to VBCurtis:

Originally Posted by VBCurtis

Those B2's are default matched up with B1 = 4e9 and up, which is quite large. I have no idea why you are quoting such B2 values with 800M!

Perhaps this quotation may be helpful:

Originally Posted by R.D. Silverman

Optimality is achieved when the algorithm spends as much time in step 2
as it does in step 1. The B2/B1 ratio depends on the relative speed
of the two steps.

So perhaps with some careful tuning of B2, -k and maybe -treefile:

1) Dedicate half of all cores to stage 1
2) Dedicate the other half to stage 2
3) ???
4) Profit!

[Madpoo]

Originally Posted by VBCurtis

Those B2's are default matched up with B1 = 4e9 and up, which is quite large. I have no idea why you are quoting such B2 values with 800M!

I have chosen to run B1 = 4.5e9 on P95 and B2 = 150e12 (k = 3, so I could go higher but expected t70 factor time rises if I do), which indeed requires 17GB for Stage 2. One thread of stage 2 per node means roughly 25 curves per day. These settings are within 2% of the best expected time for t70 (I tested 100M increments of B1 from 9e8 to 6e9). 15000 curves for a t65, 69000 for a t70.

I learned the power supplies are not strong enough for these servers- I had to turn memory down to 533 from 667 (using a "low power" setting in BIOS) and remove 8GB from two of the nodes to prevent the power supplies from shutting down when all 32 cores are in use. I also removed all-but-one hard drive from each node; my kill-a-watt suggests the 500w-rated power supplies shut down when wall plug draw exceeds 460w. Too bad they're rather non-standard supplies..

You *might* get better power savings by simply disabling turbo mode on the CPUs. In that 4-way system you have 8 CPUs and if you were able to limit each CPU by around 10-15 W, you're talking about some savings there.

Reason you might want to do that is you could get more bang for the watt by having faster memory and slightly reduced CPU speed.

I don't know the BIOS settings on that setup... on HPs it would be done with setting a power cap, among other possibilities. But something that lets you control c-states or specifically disable turbo.

[VBCurtis]

Originally Posted by Madpoo

You *might* get better power savings by simply disabling turbo mode on the CPUs. In that 4-way system you have 8 CPUs and if you were able to limit each CPU by around 10-15 W, you're talking about some savings there.

Reason you might want to do that is you could get more bang for the watt by having faster memory and slightly reduced CPU speed.

Good suggestion- unfortunately, I already did that. My only options in this basic BIOS are CPU multiplier; stock 2.5GHZ is 7.5 mult. I'm running them all at 7.0 = 2.33 Ghz, and slowing to 6 = 2.0ghz didn't help/isn't worth the lost productivity. If it didn't take 15 min to access the lower-level pair of nodes via 20ish screws, I'd pull 8GB out of those and put memory back to 667 and hope. I'll get around to that eventually..

[VBCurtis]

Originally Posted by lavalamp

I have found that GMP-ECM seems to somewhat overestimate the amount of memory it will use, so I have shied away from using -maxmem. Instead I have preferred to use -k and -treefile to manually tune it to the system I am running on.

My M1277 work has two priorities:
1. Minimum expected-time to complete a t70, as determined by the -v flag of GMP-ECM.
2. Within 2-3% of the minimum time for t70, maximize bounds to improve chances to find factors larger than 70 digits.

The RDS quote mentioned has nothing to support it, even the runs in his own paper spend 40% of stage 1 time on stage 2. I've gone a few rounds with RDS on this, and believe that GMP-ECM does not obey his theory.

As for using -maxmem, I am specifically trying to make time-efficient use of the large memory in these machines, with 6-7 of the 8 cores doing NFS work. I am not interested in dedicating more than 2 cores to ECM per machine, so there's no reason to sacrifice even 10% of performance(?). I'm trying to do something with the 32GB that I can't do with 16GB at home. I have no evidence that running enormous B2's is better than the bounds I've chosen for my first set of curves, but perhaps I should be trying to minimize t75 time since anyone with 16GB can run t70 curves relatively efficiently.

[bgbeuning]

Originally Posted by bgbeuning

If you like GPU, you may like this item on ebay

http://www.ebay.com/itm/Dell-CloudEd...AAAOSwSdZWb19I

Some more info about the c410x :
1. It requires 220VAC power.
2. The web interface to configure the iPass ports has a username / password.
There is no way to reset the password, so if you buy a used box and the
original owner changed the default (root / root) username / password then
you get to play hacker before you can configure the box.