Assessing the impact of TLB trashing on memory hard algorhitms

[Fasdurcas]

We have about 60 days for most AMD to work without an update of the ethminer. Who is responsible for update the software?

[1715536786]

1715536786

[1715536786]

1715536786

[1715536786]

1715536786

[1715536786]

1715536786

[Masked_Immortal]

is this issue just related to bandwidth, gtx970 has less bandwidth than 280x.
and what about Maxwell GPU?

[Genoil]

We have about 60 days for most AMD to work without an update of the ethminer. Who is responsible for update the software?

I filed this as a potential threat in the Ethereum bug bounty program but haven't anything from their end. Keep in mind I'm not 100% certain about this bug. It was an issue with my dagSimCL test program that may apply to ethminer as well. Unfortunately I don't have any time at the moment to further look into this. If it really is an issue, rest assured the private kernel gang has already jumped on it and resolved it, possibly using the approach that's publicly available in the dagSimCL repo. The owners of such modded kernels should be in for some serious profit...  

is this issue just related to bandwidth, gtx970 has less bandwidth than 280x.
and what about Maxwell GPU?

Maxwell cards with Compute 5.2 (GTX 9xx) only start suffering badly from TLB trashing after 2GB+ allocations, so they are fine until the switch to POS. Maxwell cards with Compute 5.0 (GTX750) have already bitten the dust and are useless for ETH mining.

Note that TLB trashing and the AMD max allocation problem are two separate issues.

[vatusasid]

The developers of Ethereum were paid 13 million Ethe. How come they could not solve this problem? The ethminer 1.0.1 is not stable yet.

[Omegasun]

Is there any news about the development of the etherminer so that it can cope with the larger DAG size. We are approaching 1280MB.

[Bagdar13]

Is there any news about the development of the etherminer so that it can cope with the larger DAG size. We are approaching 1280MB.

I have heard nothing.  I also tested this and found a problem at ~1280 as well.

I started looking on forums as I noticed a substancially drop off on the 7970 cards at ~1.2 GB.  

At this point my 280xs are down from 27 to about 24
and
At this point my 7970s are down from 22 to about 17 each (this was what supprised me and this problem is present on XFX, powercolor and one other)

This drop in performance seems to be larger than the expected drop

Oddly enough my 7870s seem to have suffered little if any performance hit and are still happy doing 15 same as at launch.

Edit my 7870s are the ghz edition, however, i also have a sapphire and a 7870 MIST (which is really a broken 7950) both of these are also unaffected.

Food for thought.

[Venon]

Is there any news about the development of the etherminer so that it can cope with the larger DAG size. We are approaching 1280MB.

I have heard nothing.  I also tested this and found a problem at ~1280 as well.

I started looking on forums as I noticed a substancially drop off on the 7970 cards at ~1.2 GB.  

At this point my 280xs are down from 27 to about 24
and
At this point my 7970s are down from 22 to about 17 each (this was what supprised me and this problem is present on XFX, powercolor and one other)

This drop in performance seems to be larger than the expected drop

Oddly enough my 7870s seem to have suffered little if any performance hit and are still happy doing 15 same as at launch.

Edit my 7870s are the ghz edition, however, i also have a sapphire and a 7870 MIST (which is really a broken 7950) both of these are also unaffected.

Food for thought.

Do you find the drop during the test or the actual mining?

[sp_]

Let the dagger grow. The ether algo will be perfect for the botnets.

[Bagdar13]

Is there any news about the development of the etherminer so that it can cope with the larger DAG size. We are approaching 1280MB.

I have heard nothing.  I also tested this and found a problem at ~1280 as well.

I started looking on forums as I noticed a substancially drop off on the 7970 cards at ~1.2 GB.  

At this point my 280xs are down from 27 to about 24
and
At this point my 7970s are down from 22 to about 17 each (this was what supprised me and this problem is present on XFX, powercolor and one other)

This drop in performance seems to be larger than the expected drop

Oddly enough my 7870s seem to have suffered little if any performance hit and are still happy doing 15 same as at launch.

Edit my 7870s are the ghz edition, however, i also have a sapphire and a 7870 MIST (which is really a broken 7950) both of these are also unaffected.

Food for thought.

Do you find the drop during the test or the actual mining?

I am now dropping in actual mining on this hardware with the dag update at block 840000; the point being is my drop in hash seems to be more than predicted by the size of DAG increase.

[Dofnatues]

Because of the drop of the hash rate. I decided to reduce the core clock frequency and keep the memory frequency the same. Is that a good idea?

[Genoil]

I just finished implementing the chunk allocation into my fork of ethminer.

https://github.com/Genoil/cpp-ethereum/tree/opencl-chunks

By allocating DAG memory in chunks (--cl-chunks <chunkSizeInMB>), issues with RAM allocation may be averted. A nice side effect of this may be (significantly) higher hashrates. Based on what I've seen from people using dagSimCL, --cl-chunks 640 yields quite good results. It may be however that there is a correlation between optimal setting of chunk size vs dag size.

I wrote this change without access to AMD hardware, so your mileage may vary. Don't bother trying this on CUDA devices, using chunks there only has a negative impact on hashrate.

  

[RustyNoman]

I just finished implementing the chunk allocation into my fork of ethminer.

https://github.com/Genoil/cpp-ethereum/tree/opencl-chunks

By allocating DAG memory in chunks (--cl-chunks <chunkSizeInMB>), issues with RAM allocation may be averted. A nice side effect of this may be (significantly) higher hashrates. Based on what I've seen from people using dagSimCL, --cl-chunks 640 yields quite good results. It may be however that there is a correlation between optimal setting of chunk size vs dag size.

I wrote this change without access to AMD hardware, so your mileage may vary. Don't bother trying this on CUDA devices, using chunks there only has a negative impact on hashrate.

  

Do you have instructions for building the problem. Do you have an exe version so that we can try.

[Genoil]

Binary is on the eth forum in mining section

[Justicemaxx]

Binary is on the eth forum in mining section

I tried with different settings of chunks, 640, 660, these figures reduce hash rate about 3 times on R280x, R290. 6x280x give about 50 MGh. At the same time setting 1300 or more does not affect the speed, the speed becomes normal, about 150 MGh, and chunks 1300 give 150 MGh. Maybe I do something wrong? ....Before starting hash, miner writes that he can't create 2 block DAG file because it is blocked GPU.

[mandica]

Binary is on the eth forum in mining section

I tried with different settings of chunks, 640, 660, these figures reduce hash rate about 3 times on R280x, R290. 6x280x give about 50 MGh. At the same time setting 1300 or more does not affect the speed, the speed becomes normal, about 150 MGh, and chunks 1300 give 150 MGh. Maybe I do something wrong? ....Before starting hash, miner writes that he can't create 2 block DAG file because it is blocked GPU.

Did your miner submit valid shares?

[Justicemaxx]

I have extracted a solo miner showed resolve, but the server node seems to be ignored, that is, the balls from the miner with chunks.

[Assanger]

I have extracted a solo miner showed resolve, but the server node seems to be ignored, that is, the balls from the miner with chunks.

Does it mean the etherminer is mining, but the shares are not recognized?

[Justicemaxx]

Yes

[Genoil]

I have extracted a solo miner showed resolve, but the server node seems to be ignored, that is, the balls from the miner with chunks.

LOL true. I'm sorry man, just knocked this out blindly without access to an actual AMD card. For now, some further testing by others have indicated there presently no need to worry about allocation problems in the near future. I wil have to verify for myself to be absolutely sure though.

[Marvell1]

I have extracted a solo miner showed resolve, but the server node seems to be ignored, that is, the balls from the miner with chunks.

LOL true. I'm sorry man, just knocked this out blindly without access to an actual AMD card. For now, some further testing by others have indicated there presently no need to worry about allocation problems in the near future. I wil have to verify for myself to be absolutely sure though.

I'd could send you one of my 7950s if you want to pay for shipping i have a bunch laying around due to no motherboards to host them in.

This dag problem is getting huge for my my 900mh/s farm is down to like 700mh/s