Bitcoin difficult still seems on the rise if anything, so I would just knock that up the variance in submitting a few shares quicker than expected in an hour, than you did last time. Variance can be quiet abit higher than predictable diff 1 shares, when doing diff 32 shares.
Variance at diff 32 is 32^2 times greater, i.e. 1024 times... so in my opinion, diff 32 is too high for current mining hardware in light of that (see my earlier post).
|
|
|
2.7.5 crashes occasionally:
[2012-08-31 03:36:40] Accepted 7de59791.15d93a38 BFL 15 pool 0 [2012-08-31 03:36:40] Accepted 2c6bbc81.dd42014b BFL 16 pool 0 [2012-08-31 03:36:40] Accepted 1d1efc9b.b8563b1f BFL 16 pool 0 [2012-08-31 03:36:40] Accepted 518e7a52.f2ccc827 BFL 20 pool 0 [2012-08-31 03:36:40] Accepted 764d3e7d.cc1a8543 BFL 20 pool 0 [2012-08-31 03:36:41] Accepted 1764a65f.7da58c1e BFL 20 pool 0 [2012-08-31 03:36:41] Accepted fc3d419d.66247269 BFL 20 pool 0 [2012-08-31 03:36:41] Accepted 781d0fcc.7e3675a1 BFL 24 pool 0 [2012-08-31 03:36:41] Accepted dc8763e5.658a1a1a BFL 24 pool 0 [2012-08-31 03:36:41] Accepted 010403c4.3e8ba7f8 BFL 30 pool 0 [2012-08-31 03:36:41] Pool 0 communication failure, caching submissionsSegmentation fault
I tried to get a backtrace of it, but when I run it in the debugger, it doesn't crash and I can't seem to run torify through gdb anyway, which is what I'm running cgminer through.
Which brings me to another request: Can you make cgminer tor friendly, so you don't have to torify it? I would love an option to be able to point cgminer to a tor server natively for all it's communications. Would save a ton of hassle.
I'm guessing you're getting the crashes once again when running through tor? I don't know the first thing about tor... yet.
|
|
|
New version: 2.7.5, 31st August 2012Survived a whole week without needing to give a hotfix update yay \o/. This release is just a minor bugfix update. I've changed my build environment so it is less work to build binaries, so I have included an fpgaonly binary as well as the full feature binary. The windows binary is now compressed with 7z since it's getting so large, and will include a different set of DLLs due to it being built differently. Human readable changelog- Hopefully fixed the bug where dynamic intensity gets stuck at -10 - Much less pool-not-providing-work fast enough messages. - Fixed share leakage to backup pools the way it was supposed to work. - Properly detect SDK2.7 and choose optimal parameters for it. - Remove the hardcoded worksize for 256 - it had the opposite effect with the new kernels - Fix the rare occasion there are many new blocks detected in a row. - Hopefully fixed the load balance and rotate strategies. Full changelog- Adjust opencl intensity when adjusting thread count to prevent it getting pegged at a value below the minimum threads possible. - miner.h max_hashes -> int64_t - Keep the local block number in the blocks structs stored and sort them by number to guarantee we delete the oldest when ageing the block struct entries. - Use correct sdk version detection for SDK 2.7 - Revert "Pick worksize 256 with Cypress if none is specified." - Test for lagging once more in queue_request to enable work to leak to backup pools. - There is no need to try to switch pools in select_pool since the current pool is actually not affected by the choice of pool to get work from. - Only clear the pool lagging flag if we're staging work faster than we're using it. - needed flag is currently always false in queue_request. Remove it for now. - thr is always NULL going into queue_request now.
|
|
|
This thread is getting Very long
Yes it is, and we shall carry on making it longer.
|
|
|
ckolivas: Only on the diff10 server or on all servers? I'm not sure why that would be, that's kind of strange.
Yep, only on the diff 10 server.
|
|
|
Well done at being the first to implement this. I quote from another thread I posted if people are trying to decide what difficulty to set: To not make variance any more painful at high hashrates, make the share return rate proportional to the square root of the hashrate instead of a constant. So a 1 GH/s miner currently returns a share every 4.2 seconds - if you make the 1GH miners difficulty 10 as a baseline, then you make 10GH miners sqrt(10) * 10 ~ diff 30. And you make 100GH miners sqrt(100) * 10 ~ diff 100.
|
|
|
If anyone wants to play around with 10Diff shares: diff10.eclipsemc.com port 8337
Give it a shot and see if anything breaks!
This is working very nicely for me and I think (variable) difficulty targets are way overdue. One thing which I have noticed, though, is that there seem to be 2 longpolls for every block change, between 15 and 20 seconds apart.
|
|
|
It's incompatible with recent versions of cgminer as it doesn't think cgminer is running and kills it off inappropriately. Try replacing the BAMT mother file in the /opt/bamt directory with this one: http://ck.kolivas.org/apps/cgminer/temp/motherNote that in the next version of bamt, apparently they're dropping support of cgminer. Ha! I was not aware of those files. Beyond my knowledge level of these things. I do not know what you modified, but it works! Sent you a small quarter of BTC tip for being so helpful! Thanks Well on the rare chance that cgminer actually *does* die, that mother file will not restart it. However it will not inappropriately kill it off as current bamt does, and I'm pretty confident about the stability of cgminer 2.7.4 anyway.
|
|
|
Lodcrappo,
Why BFGminer and not cgminer?
Better stability, better support for FPGA devices, better relationship with the developerI can't recall what I did to piss you off, but enjoy your relationship.
|
|
|
It's incompatible with recent versions of cgminer as it doesn't think cgminer is running and kills it off inappropriately. Try replacing the BAMT mother file in the /opt/bamt directory with this one: http://ck.kolivas.org/apps/cgminer/temp/motherNote that in the next version of bamt, apparently they're dropping support of cgminer.
|
|
|
Since this is one of the very few pools that has already put higher difficulty shares into trial, I quote my post from another thread that will hopefully help direct your further efforts towards this: 60 seconds per share will be way too infrequently if it's a static value. The variance will be painful. A minirig is the current highest hashrate device available and produces over 350 shares per minute. You would be setting current users to a difficulty of 350. The problem with that is the variance is the square of the difficulty so the variance will be much much higher and it would take weeks to get a stabilised return, when you are already capable of coping with hundreds of shares from the current minirig owners. To not make variance any more painful at high hashrates, make the share return rate proportional to the square root of the hashrate instead of a constant. So a 1 GH/s miner currently returns a share every 4.2 seconds - if you make the 1GH miners difficulty 10 as a baseline, then you make 10GH miners sqrt(10) * 10 ~ diff 30. And you make 100GH miners sqrt(100) * 10 ~ diff 100.
Note that the maximum efficiency of cgminer is about 70,000% with the current code, but likely limited to about half that but even then, >30,000% efficiency, combined with higher difficulty shares means most pools would cope with dramatic rises in hashrates (read ASICS) without any new protocol.
|
|
|
60 seconds per share will be way too infrequently if it's a static value. The variance will be painful. A minirig is the current highest hashrate device available and produces over 350 shares per minute. You would be setting current users to a difficulty of 350. The problem with that is the variance is the square of the difficulty so the variance will be much much higher and it would take weeks to get a stabilised return, when you are already capable of coping with hundreds of shares from the current minirig owners. To not make variance any more painful at high hashrates, make the share return rate proportional to the square root of the hashrate instead of a constant. So a 1 GH/s miner currently returns a share every 4.2 seconds - if you make the 1GH miners difficulty 10 as a baseline, then you make 10GH miners sqrt(10) * 10 ~ diff 30. And you make 100GH miners sqrt(100) * 10 ~ diff 100.
|
|
|
gpu threads 5 is not a good idea with p2pool as it takes 5 times longer to complete any work. On the other hand I doubt dropping your intensity below 8 will improve your reject rate any further either. Intensity 8 is 8.3 million hashes and at 600-700 mhash for a 7970, they're done in 1/50th of a second. That is no longer contributing to your reject rate. High rejects are the norm with p2pool, but yours is higher than normal, and perhaps it's just your p2pool setup and delays elsewhere.
|
|
|
The equation is that approximately every 4.2 GH you will find a (difficulty 1) share on average. So 1GH/s will find a share every 4.2 seconds. With 86400 seconds in a day, this equates to 86400/42 ~ 20,570 shares per day per GH/s.
|
|
|
Q: And why you write "only ignore luke-jr"?
You haven't been here very long, have you?
|
|
|
AMD is not going to make a reference model, which is why the powercolor card has been changed from a 7970x2 to a 7990.
What a bunch of dumbasses, how the hell does AMD screw the pooch like this?
Let's see... how can AMD do something stupid? Now for anyone that's been mining long enough, this question is about as rhetorical as they get...
|
|
|
I have no ztex and nothing to do with the ztex code, so I simply googled when you asked and that said -12 meant the usb version was a problem. (-12 meant device wasn't supported) It seemed that the older version of the usb library worked with more devices than the newer version ...
Maybe i must try to bild cgminer with old libusb? same version which used BTCMiner? at all possible in cgminer 2.7.4? Did you try the prebuilt binary I provide? no, i cant use prebuilt binary. I have only macs computer without OpenGl GPU. If i try start your binary, cgminer break starting with error. (need OpenGL files) I try today first time with cgminer!!!) maybe i make something wrong, maybe it can be disabled using OpenGL before starting with some options. But i cant find how? Only one solution i found: bild my own binary of cgminer.))) Oh a mac? Use libusbx, not libusb which is almost unmaintained http://libusbx.org/
|
|
|
I have no ztex and nothing to do with the ztex code, so I simply googled when you asked and that said -12 meant the usb version was a problem. (-12 meant device wasn't supported) It seemed that the older version of the usb library worked with more devices than the newer version ...
Maybe i must try to bild cgminer with old libusb? same version which used BTCMiner? at all possible in cgminer 2.7.4? Did you try the prebuilt binary I provide?
|
|
|
... If you want a pool you can rely on, miners need to accept that everyone needs to contribute, not just the forward thinking donators. Otherwise, you could mine solo I guess.
You need mining software too ......................................... Yes, I'd hope you and Con and all the team at cgminer receive coins for all your hard work too. If miners using cgminer want continued quick updates and response to changing mining conditions and protocols, they should consider an ongoing donation for the cgminer devs. However, coders can work sporadically at coding - in their spare time (what spare time?) if they have to. Graet needs to be on call (or have someone on call) 24/7, which makes a fee unavoidable. There used to be a --donation option that would submit like 0.5% or 1% of your shares to the dev's pool, but no one ever used it, so it got taken out. However, anyone who uses any software for a significant amount of time should consider donation. I used it. Thanks. It was definitely a feature that polarised the users. Some people were quite good with it. Others loathed it. Overall because I was very conscious of it not using too much power it ended up donating far less than the percentage specified, so it never really amounted to very much. Finally it became an exercise in support of its own that wasn't really donating much so I gave up and removed it. On a related note, Graet has also been the most generous donator of the pool ops to cgminer's development. Most pool ops realise that good mining software is also required, however, very very few have even donated, let alone been generous...
|
|
|
Thanks. Plan to once a stable release is out.
2.7.4 gets me ~10% less hashrate than all recent versions (including 2.7.1, 2.7.0 and 2.6.5 for fresh comparisons) on a 5870 with SDK 2.1 (ati-drivers 11.6) on Gentoo amd64. Didn't test 2.7.2 or 2.7.3. 2.7.2/3 will not even work on sdk2.1. It appears the kernel changes are not so great for sdk2.1. Oh well. There's a suggestion from IRC that those with ultra low memory clock speeds on sdk2.4/2.5 are seeing this, and going from say 150 to 250 is enough to recover it. Perhaps it's not the SDK at all but the memory speed. EDIT: Just to further explain why these changes have been made to ALL the kernels. Basically there are some small probability scenarios where some shares are missed by the existing kernel designs. It's not a lot, but given btc mining is entirely a probability game, we need everything on our side. So there's a chance that previous kernels were hashing just fine, but weren't finding shares. Will this increase U: even by a small amount? Yes, but I haven't done the math to figure out how much. Luck will make it impossible to see on a simple kernel swap since luck's effect on U is greater than the magnitude of increase this may afford.
|
|
|
|