Author Topic: New YAM version release  (Read 99663 times)

0 Members and 1 Guest are viewing this topic.

Offline sandor111

  • Newbie
  • *
  • Posts: 10
    • View Profile
Getting over 14hpm on E3-1230v2, I'm impressed!

Offline Delinquency

  • Sr. Member
  • ****
  • Posts: 370
    • View Profile
I am achieving higher speeds on my I5-2500K with some optimizations. From 5HPM to 8HPM. It starts up slow [6.8HPM] but makes its way higher.

I am now running the YAM configuration upon startup to grab as much RAM as possible from my system. This slows down applications and runs them from the HDD but it is very well worth it. (I am using four threads w/ 4096MB on a 6GB PC Setup)

I have overclocked my CPU from 3.3GHz to 4.2GHz by reapplying premium Thermal Paste and sanding the Thermal Surface of the CPU Cooler (It only has two copper pipes).

Using Process Lasso, I have decreased priorities of the running miner to run as an idle process.

« Last Edit: January 31, 2014, 07:26:04 pm by Delinquency »

Offline Delinquency

  • Sr. Member
  • ****
  • Posts: 370
    • View Profile
Please research and develop AMD processor optimization.
AMD CPU's -FX 8000 series and 9000 series:

I grab 7.5-8.5HPM with 8GB+2GB RAM and 4.5GHz OC (FX8320) still leaves GPU mining more efficient by around 50%.

It still costs 50% more per HM to capitalize but is in line with power consumption vs. GPU's. [160W+/8.5HPM{MOBO+CPU+RAM} vs. 300W+/20HPM] and [$250ish/8.5HPM for CPU+RAM vs. $400-$430 GPU/20HPM]

There are some AMD specific technologies you may be able to play around with, and please focus on AMD series as well as Intel series for your miner optimizations.

« Last Edit: January 29, 2014, 04:02:45 pm by Delinquency »

Offline h0g0f0g0

  • Jr. Member
  • **
  • Posts: 48
    • View Profile
New version of YAM Live USB miner is uploaded here:

http://memorycoin.org/downloads/livecd/releases/yam-live-1.4-M7k.zip

It fixes the issue with auto detection of a memory size. Miner is now using the maximum possible CPU threads as well as the amount of memory.

Offline yvg1900

  • Full Member
  • ***
  • Posts: 198
    • View Profile
  • BitShares: yvg1900
Yes, normally SPM converges to HPM at the end. SPM reflects actual solution finding, while HPM is pure performance metric.
Follow @yvg1900 on Twitter for yam miner updates and support

Offline bppsp1

  • Full Member
  • ***
  • Posts: 87
    • View Profile
SPM/HPM should be roughly equal at the end.

If in doubt, the most important value is SPM from what I can gather since that is what is actually being send to the pool.

Offline yvg1900

  • Full Member
  • ***
  • Posts: 198
    • View Profile
  • BitShares: yvg1900
Yes, Cfg HPM is the right metric.

Looks like you measured everything OK.

Can you please provide more details on CPU (frequency, etc) memory (# of channels and mem module characteristics) and mobo (vendor and BIOS version)?

I will look if I have similar config in my test labs.
Follow @yvg1900 on Twitter for yam miner updates and support

Offline sandor111

  • Newbie
  • *
  • Posts: 10
    • View Profile
Binding the process to each physical socket performs roughly the same as taskset

Did you measure exact HPM values? Does OS really support NUMA? What is output of

numactl --hardware

yvg1900

Yes it's enabled.

Code: [Select]
available: 2 nodes (0-1)
node 0 cpus: 0 1 2 3 4 5 12 13 14 15 16 17
node 0 size: 12278 MB
node 0 free: 123 MB
node 1 cpus: 6 7 8 9 10 11 18 19 20 21 22 23
node 1 size: 12288 MB
node 1 free: 990 MB
node distances:
node   0   1
  0:  10  20
  1:  20  10
Code: [Select]
MMC Agg. SPM: 12.304, HPM: 12.943; Rnds C/I: 9919/2596, Don. C/I: 109/22; Cfg/Wkr SPM: 12.304/0.5804, Cfg/Wkr HPM: 12.945/0.5884 10028 rnds AV=1, ART=101969
With numactl, roughly 6.45 hpm per socket, so 12.9 hpm total.
How do I measure the exact hpm, are the reported values accurate?

Offline volkantipi

  • Newbie
  • *
  • Posts: 10
    • View Profile
Hi, Xeon E5-2609 4 GB ram what is the best config?

Offline luckye

  • Newbie
  • *
  • Posts: 19
    • View Profile
I think that the New k version works better that the previous one because is more flexible...for a system like the mine with a lot of physical and virtual cores, the memory size is crucial.
I'm getting 2.4 hpm more !

Offline bppsp1

  • Full Member
  • ***
  • Posts: 87
    • View Profile
I take back what I said earlier yvg1900

Figured out had random other issue causing issues with running yam instance.

Right now running av=1 and m=1024*threads resulting in 3-4% performance gain over previous version.

Thanks

Offline yvg1900

  • Full Member
  • ***
  • Posts: 198
    • View Profile
  • BitShares: yvg1900
Binding the process to each physical socket performs roughly the same as taskset

Did you measure exact HPM values? Does OS really support NUMA? What is output of

numactl --hardware

yvg1900
Follow @yvg1900 on Twitter for yam miner updates and support

Offline h0g0f0g0

  • Jr. Member
  • **
  • Posts: 48
    • View Profile

Offline yvg1900

  • Full Member
  • ***
  • Posts: 198
    • View Profile
  • BitShares: yvg1900
SPM is a metric (Solutions Per Minute) that depends very much on luck. Brute performance metric is HPM.

Fact that you see much smaller ART means that many threads share same worker memory area.

Can you post here output from the miner where it shows Threads/Workers and memory size in use?

yvg1900
Follow @yvg1900 on Twitter for yam miner updates and support

Offline bppsp1

  • Full Member
  • ***
  • Posts: 87
    • View Profile
I notice that the new version doesnt have a constant 99.9-100% peg on each thread, it fluctuates alot between 92-98 on each threads and rarely pegs at 99.9-100%

Is this by design, surely such fluctuations would kill extra hashrate ?

That would be the case if you use m=1024 and many threads. In this case mining will suffer from inter-thread sync contention.

Max perf can be obtained when m=1024*numberOfThreads, which is exactly the config from prev version.

Tried both, unfortunately the only thing better for me using version k is the much lower ART from 60000 down to 2000 , SPM/SPM suffered about 2-3% over long running session average.