archive snowblossom slack general
prev 2018-08-24 next
2018-08-24 02:05:39
PAPACABEZA
@Fireduck don't stop with FrostyTrader --- where there's innovation, there's a way, don't get scared off by liability. We should talk through and find a way to address it but it probably is in some way.
2018-08-24 02:06:59
Fireduck
I'm not concerned about civil liability.
2018-08-24 02:07:54
Fireduck
It sounds like qtrade is about a month out so makes sense to wait for them
2018-08-24 03:12:52
alistar
integrated with qtrade？
2018-08-24 03:37:18
Fireduck
http://qtrade.io is an exchange that is working on listing snowblossom
2018-08-24 03:38:01
penfold
@Fireduck the slowdown if the whole field is not available… you had a neat formula… something containing (1/2)^6?
2018-08-24 03:38:11
penfold
ran out of scrollback in Slack I think :slightly_smiling_face:
2018-08-24 03:38:38
penfold
well 1/2 assuming 50% available I guess
2018-08-24 03:40:22
Fireduck
yeah, just basic probabilities
2018-08-24 03:42:33
penfold
so is that right, if you had 50% of the field the effective hashrate would be reduced to ~1.56% ?
2018-08-24 03:44:13
penfold
I think I’m getting about 78% efficiency with 128GB RAM, around 122-123 is used for caching the field, so ~95% of the field… ouch :slightly_smiling_face:
2018-08-24 03:46:32
penfold
I guess I need to try Arktika or w/ever it is to fill out that last 5%
2018-08-24 03:52:43
Fireduck
if you have some percent of the field p, then it is 1/(p^6) is the prob of being able to check a complete pow
2018-08-24 03:53:46
Fireduck
nevermind the 1 over, losing my mind
2018-08-24 03:54:22
Fireduck
anyways, the problem with PoolMiner is that it doesn't discard when it gets a cache miss, it goes to disk and gets it
2018-08-24 03:54:28
Fireduck
which doesn't help you much
2018-08-24 03:54:46
Fireduck
Arktika can discard it (or more correctly, put it on a layer with no or few assigned threads)
2018-08-24 03:58:11
Fireduck
Also, with Arktika you could host those last 6 GB on a separate machine and have them use each other
2018-08-24 06:31:46
penfold
I guess the idle GPU with 8GB+ could work too
2018-08-24 06:32:12
penfold
anyone tried GPU yet?
2018-08-24 06:32:30
penfold
6x skein is going to be slow though, I presume the slowdown is too much given the limited RAM
2018-08-24 10:13:11
mjay
You´d need to connect enough GPU´s to get at least close to 128GB, otherwise (p^6) will kill your hashrate
2018-08-24 10:14:36
mjay
I did a few experiments on this, not actual GPU, but just loading the needed data from ram without any hashing, just like it would have to be done when offloading the hashes to gpu. This was only slightly faster than regular CPU hashing
2018-08-24 10:15:56
mjay
I guess its still the memory bandwidth thats limiting here. Had 100% cpu load on my machine, but it actually was the cpu waiting for the memory to respond
2018-08-24 10:17:09
mjay
Copying all data to the GPU would need even more I/O. I don´t think this is faster.
2018-08-24 10:17:32
mjay
Unless someone can get their hands on a 128GB+ GPU. Is there even such a thing?
2018-08-24 12:42:55
penfold
interesting stuff, and no I don’t think as of now there are GPUs with >32GB
2018-08-24 12:45:14
mjay
Even 32GB are way too expensive, like the V100
2018-08-24 12:46:49
Fireduck
plenty of bandwidth if you batch it like arktika does for network access
2018-08-24 12:47:54
Fireduck
but that would be some fancy gpu programming
2018-08-24 12:51:14
penfold
given PCI-e x16 is ~16GB/s it’d just be very expensive RAM methinks :slightly_smiling_face:
2018-08-24 12:56:56
Fireduck
can do a pretty insane hash rate with that
2018-08-24 12:57:27
penfold
oh? isn’t RAM faster? I thought I was seeing ~30GB/s on DDR4
2018-08-24 12:58:03
Fireduck
yeah, but you are absolutely cpu bound way before that
2018-08-24 12:58:21
penfold
ohh right, hmm, that’d be a nice stat to output in the miner
2018-08-24 12:58:37
penfold
what is it called? memory bandwidth? or more genreally, I/O bandwidth?
2018-08-24 12:58:59
Fireduck
there is a fake mode in arktika to see what the cpu limit is
2018-08-24 12:59:10
penfold
I guess I really need to download arktika :slightly_smiling_face:
2018-08-24 12:59:28
penfold
not sure where I got 30GB/s from I only get 7-8GB/s
2018-08-24 13:00:12
Fireduck
I think I get 24 GB/s on a gce instance
2018-08-24 13:00:13
penfold
well, that’s using `mbw`, whatever that measures :smile:
2018-08-24 13:00:19
Fireduck
but not sure
2018-08-24 13:00:21
penfold
oh maybe that’s what it was, I did test on GCE
2018-08-24 13:08:49
mjay
do you also get 24GB/s with 16-byte-random-access?
2018-08-24 13:08:57
mjay
for me these numbers are a lot lower
2018-08-24 13:10:42
penfold
no but I’d like to know what those numbers are… any tools that can do it? or had to roll your own?
2018-08-24 13:11:29
penfold
I think the peak was large block memcpy using `mbw`
2018-08-24 13:11:56
mjay
I just modified the java miner, about 500MB/s on a 40 core 4 socket server
2018-08-24 13:12:08
mjay
this is just counting the 16 bytes each
2018-08-24 13:12:22
mjay
18GB/s with memcpy otherwise
2018-08-24 13:12:34
penfold
wow, quite stark
2018-08-24 13:12:47
penfold
would that have two (or more) memory channels per socket?
2018-08-24 13:13:30
penfold
I wonder, is it possible to calculate effective I/O bandwidth if you know the hashrate?
2018-08-24 13:13:53
penfold
I suppose it’s just 16*6*hashrate
2018-08-24 13:14:43
penfold
meaning I’m getting about 50MB/s but <80% efficiency (not quite enough RAM) so maybe ~64MB/s is possible for my hardware
2018-08-24 13:15:43
penfold
so maxing out PCI-e x16 would be a 250x improvement
2018-08-24 13:16:46
penfold
to get enough memory though I’d need to drop some to x8 and/or x4
2018-08-24 13:16:58
penfold
or, limit to 2-3 GPUs per motherboard I suppose
2018-08-24 13:17:03
penfold
gets expensive quickly
2018-08-24 13:20:10
mjay
This server has 2 memory channels per CPU, thats 8 total. DDR4
2018-08-24 13:20:24
penfold
nice
2018-08-24 13:21:21
mjay
With server hardware you can easily connect 8 GPUs, each 16x
2018-08-24 13:21:39
mjay
this gets expensive, yes :wink:
2018-08-24 13:21:51
penfold
laugh, yep, plus I’m not keen on having a jet engine in the same room as me :slightly_smiling_face:
2018-08-24 13:22:17
mjay
You dont sleep next to these things
2018-08-24 13:22:23
mjay
or: you cant.
2018-08-24 13:23:06
penfold
haha, yes… I don’t intend to but the thing is, using a couple of appliances + a rig in one part of the house will trip the breaker… so I run it in part of the house that while in a separate room, is still audible at night with the door open… thin walls X)
2018-08-24 13:23:42
mjay
glad I have a cellar to do this
2018-08-24 13:23:53
mjay
big walls, several doors, always cool
2018-08-24 13:24:11
mjay
also very humid, but hey :smile:
2018-08-24 13:24:33
penfold
maybe the hardware will be obsolete before it corrodes too badly :slightly_smiling_face:
2018-08-24 14:16:08
Fireduck
keep it hot and the relative humidity is lower
2018-08-24 17:29:18
Rotonen
@mjay one would have to do something silly like writing an in-gpu nvme direct driver, but that’d still not be significantly quick
2018-08-24 17:30:07
Fireduck
@Rotonen Imagine the GPU working on a bunch of PoW units and filling a buffer with the offsets of the words it needs from the snowfield.
2018-08-24 17:30:32
Fireduck
Then the CPU side reads that buffer in chunks and fills in those words in another buffer and sends the results back.
2018-08-24 17:30:53
Fireduck
Basically can offload the hash processing to the GPU and make use of the CPUs memory bus bandwidth.
2018-08-24 17:31:11
Rotonen
i have imagined this
2018-08-24 17:31:19
Rotonen
it has not turned me on
2018-08-24 17:31:32
Fireduck
haha, then what do you hang around here for? :wink:
2018-08-24 17:31:53
Rotonen
8TB fields
2018-08-24 17:32:19
Fireduck
ok, good
2018-08-24 17:32:24
Fireduck
I have the 4tb field generated
2018-08-24 17:33:53
Rotonen
@mjay if you squeeze more than 300kH/s per memory channel from RAM, do tell
2018-08-24 17:34:23
Rotonen
high end NVMe goes up to close to 200kH/s per disk
2018-08-24 17:34:43
Rotonen
-> for now AMD EPYC is the thing
2018-08-24 17:48:54
Clueless
@Fireduck I'm prepared to seed a lot more
2018-08-24 17:49:36
Fireduck
make sure you have the things there, especially chunked for 7,8,9
2018-08-24 17:54:23
Clueless
@Fireduck I haven't setup any of the chucked ones yet, nor monitors for them. I may have to buy some storage if we count those. Which is fine. I've been looking for an excuse to buy more drives. It'll triple my storage.
2018-08-24 18:17:58
Fireduck
I don't think your monitors exist
2018-08-24 18:19:27
Fireduck
gradle gradle gradle, I made you out of clay
2018-08-24 18:19:35
Fireduck
Why are there so many terrible build systems
2018-08-24 19:43:08
THX 1138 4EB
Would you expect the following to make a shortcut that will open node.sh?
2018-08-24 19:43:12
THX 1138 4EB
[Desktop Entry] Encoding=UTF-8 Type=Application Name=Snowblossom node Name[en_GB]=Snowblossom node Icon=/home/thx11384eb/Downloads/snowflake_logo.png Exec=/home/thx11384eb/snowblossom-1.3.0/node.sh Comment[en_GB]= Terminal=true StartupNotify=true X-KeepTerminal=true
2018-08-24 19:44:08
THX 1138 4EB
When I click on the shortcut, the cursor gets a sand timer but no terminal window opens
2018-08-24 20:29:13
Fireduck
I have no idea what that configuration notation is. Some sort of window manager specific thing? It looks reasonable. The only thing suspect is that node.sh assumes you are in working directory that node.sh is in. Maybe specify a working directory or put a 'cd' into the start of node.sh ?
2018-08-24 21:21:36
Rotonen
that's windows `.lnk`
2018-08-24 21:21:41
Rotonen
i guess
2018-08-24 21:23:26
Fireduck
yeah, but it is a linux path
2018-08-24 21:25:26
Rotonen
gnome?
2018-08-24 21:28:08
Rotonen
`https://<redacted>/video/search?search=gnome`
2018-08-24 21:34:12
Fireduck
The movie I linked is not worth searching for, watching or buying
2018-08-24 22:30:02
Clueless
@THX 1138 4EB It should work, but you'd have to install snowblossom as your own user.
2018-08-24 23:20:39
Clueless
@Fireduck @THX 1138 4EB So typically the scripts I provided defaulted/leaned towards creating a dedicated system user/service. This favors servers and such... but.. I should probably improve upon the two avenues of install. Headless, and headed. So gui/windows/ubuntu users can point and click, and more easily run as a GUI application.
2018-08-24 23:21:08
Clueless
I think I'll do that built into my Python GUI run.
2018-08-24 23:28:58
Rotonen
@Clueless systemd has user services
2018-08-24 23:59:55
Fireduck
I usually run new suspicious crypto coins in a terminal under screen