archive snowblossom slack general
prev 2018-10-29 next
2018-10-29 00:00:04
mjay
why is it silly?
2018-10-29 00:00:22
Rotonen
and someone should do the mathemathics on what's the collision chance of shuffling the fields in regards to getting a higher ratio of nicer seeks
2018-10-29 00:01:22
mjay
why should shuffling improve it?
2018-10-29 00:01:45
Rotonen
128bit is just not a nice thing to seek in as that'd require page size tweaks
2018-10-29 00:03:09
mjay
it should just proof you did something, it could do more in the same time, but thats not the point
2018-10-29 00:03:15
Rotonen
well, playing around with the chance of 256 reads happening from the same 4k block
2018-10-29 00:03:33
Rotonen
but that's quite on the fringe of maybes
2018-10-29 00:04:20
Rotonen
but yeah, 16 byte reads off a gpu, do provide the work for reverification if you tackle that
2018-10-29 00:04:58
Rotonen
heh, nvidia has blogged close enough https://devblogs.nvidia.com/maximizing-unified-memory-performance-cuda/ Detailed explanation of CUDA Unified Memory page migration to improve your understanding and help you optimize and get the best Unified Memory performance.
2018-10-29 00:06:12
Rotonen
`Considering that Unified Memory introduces a complex page fault handling mechanism, the on-demand streaming Unified Memory performance is quite reasonable. Still it’s almost 2x slower (5.4GB/s) than prefetching (10.9GB/s) or explicit memory copy (11.4GB/s) for PCIe.` gives you a ball park
2018-10-29 00:06:23
mjay
okay, first test, directly accessing 16 byte chunks from global memory, using openCL, 18.5M accesses/s on a 1080TI
2018-10-29 00:06:40
Rotonen
but they did tiered caching and all the trimmings for you
2018-10-29 00:06:49
mjay
offloading will be faster, thats the next test
2018-10-29 00:07:32
Fireduck
there is a fun way around the 4k read problem
2018-10-29 00:09:36
Rotonen
that caps out at 40GE for anything consumer obtainable
2018-10-29 00:09:49
Rotonen
though not many even have 10GE
2018-10-29 00:10:41
Rotonen
@Fireduck or are you saying that'd be faster over a switch with two network interfaces on the same computer? :smile:
2018-10-29 00:11:16
mjay
Better use localhost :smile:
2018-10-29 00:11:30
Rotonen
loopbacks are unix sockets under the hood and thus 4k? :stuck_out_tongue:
2018-10-29 00:12:00
Rotonen
i'm trying to wrap my head around there which part is able to feed faster, as there's ultimately down the line a 4k read system somewhere down the line
2018-10-29 00:12:03
mjay
Is every 16 byte chunk sent on its own???
2018-10-29 00:12:20
mjay
That would be a waste
2018-10-29 00:12:24
Rotonen
on the network level, yeah, can be, but what feeds into that and how
2018-10-29 00:12:48
Rotonen
meh, only like 20% overhead all inclusive? :stuck_out_tongue:
2018-10-29 00:12:56
mjay
I´d fit at least 50 of them into one packet
2018-10-29 00:13:19
mjay
packets have a header, too
2018-10-29 00:13:32
mjay
more than 16 bytes total
2018-10-29 00:18:02
Rotonen
@mjay so seems cuda actually has everything one would need and no need to orchestrate if you think of each execution unit as just a cpu doing one fetch, and all the layers within the gpu as sacrificial caches for the main system memory, and give priority to the gpu - all the gpu internal 'cache hits' would be an edge
2018-10-29 00:20:06
mjay
Of cause. But right now with 128GB field, and most consumer GPUs way below 16GB, thats not even 15%
2018-10-29 00:20:31
mjay
once field 8 activates even less
2018-10-29 00:20:43
Rotonen
when mining off an ssd, everything you get on top counts
2018-10-29 00:21:00
Rotonen
if hardware you might already have doubles your hashrate, win
2018-10-29 00:21:23
Rotonen
though nonsense in regards to hashes per watt
2018-10-29 00:21:47
mjay
it could stay in its lower power stages
2018-10-29 00:22:12
mjay
compared to an ssd its still way faster
2018-10-29 00:22:54
Rotonen
oh, that's the kind of fun for which one can get paid in fiat - how to idle GPGPUs in flight
2018-10-29 00:24:01
Fireduck
The idea is 4k to CPU, then marshalled over network. That way you use multiple machines to access a cluster of memory machine
2018-10-29 00:24:33
Fireduck
In practice it works great
2018-10-29 00:24:56
Fireduck
My r900 can do about 300kh/s on it's own
2018-10-29 00:25:19
Fireduck
But enabled about 1mh/s by sharing over 1g network
2018-10-29 00:25:34
Fireduck
To other machines that Max their CPU as well
2018-10-29 00:27:20
mjay
What we are aiming for is an arktika-gpu-client
2018-10-29 00:28:15
mjay
that uses the snowfield either on a local machine or a network source
2018-10-29 00:30:36
Rotonen
i'm just emptying the dishwasher and shooting the breeze
2018-10-29 00:37:01
mjay
around 22Mio. accesses per second seems to be about max for a 1080TI
2018-10-29 00:37:14
mjay
16 bytes, random position, 16 byte aligned, from 10GB of data
2018-10-29 00:38:11
mjay
power draw <70W
2018-10-29 00:50:49
Rotonen
that'd be like 2,5MH/s over that plus penalties from misses, so currently that'd land you ~200kH/s extra? that's not bad per watt
2018-10-29 00:51:26
Rotonen
i'm rounding down and probably not enough, but ballpark-wise something like 1..2 extra SSDs for hardware you might already have?
2018-10-29 00:51:50
Rotonen
assuming you have the 10GB of RAM to dedicate to that, otherwise extra penalties and no sense?
2018-10-29 00:52:58
Rotonen
so someone could have a cheap longterm viable miner off an SSD and have the GPU profit switch between boosting snow and doing whatever else, i like the narrative - now just someone would have to put a lot of effort into software to try that out
2018-10-29 00:53:58
Rotonen
probably a ~80h project for someone who has done something similar before? (80h for a PoC implementation is a lot of work for something this specific, mind you)
2018-10-29 00:54:13
Rotonen
@Fireduck did you even spend that much on snowblossom so far? :smile:
2018-10-29 00:56:22
Rotonen
and about to hit 1BTC on 24h trading volume on qtrade too
2018-10-29 00:57:50
Fireduck
I've probably spent 200h on snowblossom
2018-10-29 00:57:56
Fireduck
Hard to estimate
2018-10-29 00:58:21
Fireduck
More than that, I was effectively full time for six weeks
2018-10-29 00:59:53
Rotonen
ah you had some luxury between jobs or how'd you pull that off?
2018-10-29 01:00:17
Rotonen
no wonder most of the hard stuff actually works
2018-10-29 01:08:28
Fireduck
Yeah, took time between jobs
2018-10-29 01:08:48
Fireduck
Quit Google and started at Axon
2018-10-29 03:00:13
Fireduck
I still wonder who the hell is mining with field 9 and why.
2018-10-29 03:00:24
Fireduck
maybe they just have 1tb SSDs and why not
2018-10-29 03:01:23
quantumblockz
There is not benefit to using it right now. Even downloading it would be a pain. I imagine the person using it doesn't know the correct snowfield
2018-10-29 03:02:58
Fireduck
I kinda doubt that
2018-10-29 03:03:04
Fireduck
it is certainly possible of course
2018-10-29 04:34:34
alistar
when field 9?
2018-10-29 04:34:44
alistar
3G harsh rate?
2018-10-29 07:08:37
izzysurkin
as people are too lazy to just check the block explorer, maybe worth placing current active field alongside some other stats on the front page of main site @Fireduck
2018-10-29 09:22:17
mjay
also, maybe tell people on the snowfield-torrent-page that the fields < 7 are not needed anymore
2018-10-29 09:22:44
mjay
I know at least 2 people who downloaded them
2018-10-29 09:58:33
Rotonen
@alistar 8 at ~4G, 9 at ~8G
2018-10-29 10:02:44
mjay
2^(25+2*field)/600
2018-10-29 10:05:07
mjay
9 is at 14,6G
2018-10-29 10:05:28
Rotonen
thought it just doubles, nice
2018-10-29 10:06:20
Rotonen
maybe those should go onto the snowfields explanation page
2018-10-29 10:06:54
Rotonen
and also say they’re all the same field, with ’including up to’
2018-10-29 10:11:16
Rotonen
for anyone lazy, see the values table http://m.wolframalpha.com/input/?i=2%5E%2825%2B2*n%29%2F600 Wolfram|Alpha brings expert-level knowledge and capabilities to the broadest possible range of people—spanning all professions and education levels.
2018-10-29 11:16:15
izzysurkin
@mjay when I first came across snow I learnt this hard lesson too as I aggressively skimmed the docs due to wanting to mine some snow asap. little patience goes a long way
2018-10-29 11:17:20
mjay
Its probably best to download field 8 right now so you are ready when it switches
2018-10-29 13:10:37
izzysurkin
@mjay already on it :smile:
2018-10-29 14:03:28
lajot
i shouldnt be able to edit that right?:grin:
2018-10-29 14:03:45
Fireduck
@Clueless can you update your testnet node to the latest in git?
2018-10-29 14:03:49
Clueless
sure
2018-10-29 14:04:03
Clueless
@lajot are you saying it'll let you edit it?
2018-10-29 14:04:17
lajot
i think so yes
2018-10-29 14:04:24
lajot
@lajot set the channel topic: WIKI https://github.com/snowblossomcoin/snowblossom/wiki hello
2018-10-29 14:04:57
Clueless
ah, I thought you meant the wiki, yeah, you shouldn't be able to change that either, thanks. :P
2018-10-29 14:06:17
Clueless
@Fireduck I thought I was current already
2018-10-29 14:07:19
Clueless
`Oct 29 07:07:08 snownode node-testnet.sh[954]: [2018-10-29 07:07:08] INFO snowblossom.node.SnowBlossomNode <init> Starting SnowBlossomNode version 1.4.0`
2018-10-29 14:09:34
Fireduck
1.4.0 is ages ago
2018-10-29 14:09:42
Fireduck
you want 1.4.1-dev
2018-10-29 14:18:40
Clueless
herk
2018-10-29 14:20:14
Clueless
ah, an update script broke
2018-10-29 14:23:44
Fireduck
I made a change such that low fee transactions only get 100k from each block
2018-10-29 14:23:53
Fireduck
that way old clients that don't do the fee right will still work
2018-10-29 14:24:02
Fireduck
but it won't be possible to fill all the blocks for free anymore
2018-10-29 14:28:57
Clueless
yeah, just to warn you all, I'm redoing my monitoring not happy with my old monitoring setup, so some things like this may slip under the radar
2018-10-29 15:07:21
mjay
@Fireduck do all mining nodes have to upgrade or will the network still accept blocks with more than 100k low fee transactions?
2018-10-29 15:15:15
Fireduck
it is entirely block creator deciding, other nodes will accept the blocks just fine
2018-10-29 15:15:25
Fireduck
not a protocol change
2018-10-29 15:15:36
mjay
okay, great
2018-10-29 19:33:33
Fireduck
I am talking with a guy on reddit, can anyone provide any input?
2018-10-29 19:33:34
Fireduck
https://www.reddit.com/r/snowblossom/comments/94kuwt/xmss_vs_rsa_large_key/e8ocv1z I appreciate you taking the time to set me straight. I am an expert in blockchain cryptocurrency and pretty solid on traditional crypto but the...
2018-10-29 19:33:43
Fireduck
Specifical, WTF stateless signing?
2018-10-29 19:34:05
Fireduck
also, why the hell does quantum resistant mean the same thing as quantum safe?
2018-10-29 19:34:58
Fireduck
I am pleased when anyone looks hard enough to tell me I am wrong
2018-10-29 20:02:22
mjay
As far as I know algorithms are only quantum safe if there is provably no quantum algorithm with a sub-exponential number of qubits, where quantum resistant is just we don´t know any algorithm
2018-10-29 20:04:41
Fireduck
ah, ok, that is the distinction I was missing
2018-10-29 21:57:59
Gamaranto
Any project admin/dev I can talk to? :smile:
2018-10-29 22:01:15
Clueless
@Gamaranto what's up? :P
2018-10-29 22:11:05
Rotonen
@Fireduck why i keep telling fighting unknown unknowns in a practical application is a bit silly
2018-10-29 22:12:33
Rotonen
@Fireduck someone eventually needs to hilbert up all the idea spaces, but that’s meta engineering
2018-10-29 22:18:10
Fireduck
@Gamaranto I am here too.