2018-09-11 00:13:38
apparently something is still pointing people towards it

Rotonen
anyone with a dual 6276?
2018-09-11 22:03:59
thats 2 memory channels per socket?
http://www.cpu-world.com/CPUs/Bulldozer/AMD-Opteron%206276.html
if that's the case i'd guess 800kH/s .. 900 kH/s, but am more than welcome to be corrected

Rotonen
4 channels i think
2x2
but i may be wrong
The number of controllers: 2
Memory channels per controller: 2
@Rotonen if memory is "reserved" can we get more hasharte?
or about the same
2018-09-11 22:09:51
that'd only be relevant with very asymmetric NUMA nodes and very large snowfields, but yes, but not yet

Rotonen
2018-09-11 22:10:14
think of something like a 2TB field on a multisocket AMD EPYC, that and up

Rotonen
i rented a 6276 x2
2018-09-11 22:10:37
i have some sketches, i welcome all updates to them
https://snowplough.kekku.li/wizard/ A quickstart wizard to help you get started with mining Snowblossom.

Rotonen
just some hours ago
2018-09-11 22:10:54
click on links in the table to navigate around topics

Rotonen
Im aware of your pool / website
2018-09-11 22:11:09
but as you have 4 memory channels, i'd estimate 800kH/s

Rotonen
so 1.6 mh?
2018-09-11 22:12:02
i'm not sure how i'm reading that processor spec, 0,8 or 1,6, but yes

Rotonen
2018-09-11 22:13:01
oh that's a multipackage thingy like the core2quad was, kk, so 1,6MH/s is plausible

Rotonen
maybe if i use pre cache field
for using cpus cache
i get more?
2018-09-11 22:13:34
no, memfield

Rotonen
as these cpus have a lot of cache
2018-09-11 22:13:47
no, that'll happen below the operating system

Rotonen
2018-09-11 22:14:08
completely automatic, not something you can fiddle with unless you write your own operating system

Rotonen
ok tell me another think
gh
128gb is not enough
correct?
2018-09-11 22:14:50
yep, but you can still probably get 1MH/s off it

Rotonen
2018-09-11 22:15:05
just don't precache, the filesystem cache will eventually have enough of it in memory

Rotonen
i tried a vps with 300 gb
2018-09-11 22:15:26
alternatively you can precache and tweak the abort settings for how manieth can be a memory cache miss for it to give up or not

Rotonen
was getting 1.6mh
but its shared memory
2018-09-11 22:15:55
with virtual machines it is very hard to judge how many memory channels your seeks get mapped over

Rotonen
exactly
2018-09-11 22:16:42
but it's reasonable to assume they pack 1 .. 2 socket blades into 1U units as their backing hardware, and then overprovision those for their margin

Rotonen
2018-09-11 22:17:16
the vendors probably have conference talks on that kinda stuff on youtube if someone is actually interested in trying to reverse engineer how to max that out

Rotonen
2018-09-11 22:18:03
Soon they will make snowblossom optimized hardware

Fireduck
2018-09-11 22:18:13
aka file server

Rotonen
2018-09-11 22:18:54
i'm pretty sure pixar and other similar houses would have just the right stuff for snowblossom

Rotonen
2018-09-11 22:21:05
Probably

Fireduck
when is the private miner released?
2018-09-11 22:33:16
there is a private miner?

mjay
ofc there is one
such big hashrates
2018-09-11 22:35:59
hmm

mjay
2018-09-11 22:36:03
do you think its gpu?

mjay
2018-09-11 22:36:05
that's not a software problem, though

Rotonen
yes
i do think its gpu
along system ram
2018-09-11 22:37:37
how'd that even work?

Rotonen
2018-09-11 22:38:13
as you have 16GB per card on high end datacentre gear, there'd still be ram in the mix, limiting the seek speeds?

Rotonen
2018-09-11 22:38:23
gpu memory access is fast, but you need 128GB

mjay
2018-09-11 22:38:28
ref. how that works for not enough ram vs. disk

Rotonen
2018-09-11 22:38:43
The V100 is also in 32GB, but at a very high price

mjay
2018-09-11 22:39:00
well, those cost along the lines of 15k usd a pop

Rotonen
2018-09-11 22:39:11
So .. 4xV100 32GB + NVLink, maybe possible

mjay
2018-09-11 22:39:19
nvlink is not that fast

Rotonen
couldnt the gpu act
as cpu
2018-09-11 22:39:47
you still need 128GB gpu memory

mjay
2018-09-11 22:39:58
i'll rather believe someone is throwing a mainframe at this before believing someone solved the interconnect problem, which is a pretty hard HPC problem

Rotonen
i said act as cpu
gpu*
meaning
you have x cuda cores
2018-09-11 22:40:11
the number crunching is not limiting the hash rate

Rotonen
plus system memory
2018-09-11 22:40:31
even CPUs could do more hashes, you just cannot get the data to them fast enough

Rotonen
2018-09-11 22:40:55
I experimented with this quite a lot. The point is - the gpu needs a lot of data from memory when mining. So much its saturating the PCIe-Bus

mjay
2018-09-11 22:41:21
While the memory is fully loaded

mjay
mjay may you share your flags in private?
as we are still early miners
:wink:
2018-09-11 22:41:37
and even if you have an omnipath fabric or equivalent, that'll still be about as fast as ram is, or slower

Rotonen
2018-09-11 22:41:41
what flags?

mjay
for mining
2018-09-11 22:42:01
its more about cheap hardware

mjay
ram its not cheap
2018-09-11 22:42:15
DDR3 Reg is

mjay
hmm how much do you think 1 snow is worth
if we started now
im looking at market and hashrates
and i dont get how ones get 5+ mh
2018-09-11 22:43:05
@dk3 but if you're interested in figuring this stuff out, start reading into HPC interconnects, they're not magic (also limiting AI research currently, so there is money being thrown at solving those issues)
https://www.intel.com/content/www/us/en/high-performance-computing-fabrics/omni-path-driving-exascale-computing.html Find the latest benefits on Intel® Omni-Path Architecture and how Intel is clearing the path to Exascale computing, addressing tomorrow’s HPC issues.

Rotonen
@Rotonen maybe we just buy an movidius acelarator
2018-09-11 22:43:37
do note the pools just sum the hash rates of multiple miners if they mine to the same address

Rotonen
instead of a cpu
:wink:
i know how it works @Rotonen just trying to figure this out as a miner
2018-09-11 22:44:08
I also tried a xeon Phi, which is actually slower than a cpu

mjay
i think its more about cores
tbh
*threads
2018-09-11 22:44:37
heh, was not expecting anyone to have actually tried the knights landing stuff out

Rotonen
2018-09-11 22:44:53
once the memory channels are fully loaded more threads are useless

mjay
2018-09-11 22:44:54
(and descedants of the mad 'let us superglue 100 pentiums together' project)

Rotonen
2018-09-11 22:45:24
@mjay say, would you have any IA-64 laying around?

Rotonen
2018-09-11 22:45:59
@Rotonen Knights corner, I cheaped out on this one :wink:

mjay
2018-09-11 22:46:28
No never worked with IA-64

mjay
2018-09-11 22:47:08
afaik at least huawei is still selling them

Rotonen
2018-09-11 22:47:16
i think fujitsu finally gave up

Rotonen
2018-09-11 22:47:52
do they have advantages when it comes to mining snow?

mjay
2018-09-11 22:49:05
they have a set of caching choices i've not had access to, probably not, but i'm curious

Rotonen
2018-09-11 22:49:21
and they updated the itanium line with the 9700 with just last year

Rotonen
2018-09-11 22:50:08
disabling caching would be the best choice?

mjay
2018-09-11 22:50:38
well, they do clever stuff to avoid the cycle miss

Rotonen
2018-09-11 22:50:44
or, so they say on paper

Rotonen
2018-09-11 22:51:30
Is there no equivalent on x86?

mjay
2018-09-11 22:52:44
iirc they do hypertheading in some weird registry magic way instead of just cache interleaving

Rotonen
2018-09-11 22:52:53
not finding the actual chart i remember, but
https://en.wikipedia.org/wiki/Montecito_(processor) Montecito is the code-name of a major release of Intel's Itanium 2 Processor Family (IPF), which implements the Intel Itanium architecture on a dual-core processor. It was officially launched by Intel on July 18, 2006 as the "Dual-Core Intel Itanium 2 processor". According to Intel, Montecito doubles performance versus the previous, single-core Itanium 2 processor, and reduces power consumption by about 20%.[1] It also adds multi-threading capabilities (two threads per core), a greatly expanded cache subsystem (12 MB per core), and silicon support for virtualization.

Rotonen
2018-09-11 22:53:19
When I look at the kernel source of madvise, MADV_RANDOM would be the flag to set

mjay
when we reach 256 snowfield
will we see an hash srop?
drop?
2018-09-11 22:53:49
average diff 41, probably 4G nethash (minus diff creep per block)

Rotonen
ok my question is
2018-09-11 22:54:04
I see special code for alpha, mips, parsic and xtensa

mjay
or should be can snowfields go back?
2018-09-11 22:54:08
and yes, that's been customary so far, but big miners usually have the next field anyway

Rotonen
2018-09-11 22:54:22
snowfields never go back, the current one activated at diff 39

Rotonen
ok got it
2018-09-11 22:54:53
which sorta drove away a lot of the early bigger miners as memory on the cloud is expensive

Rotonen
it is
2018-09-11 22:55:26
the best thing i see happening is people mining at multiples of 100kH/s on NVMe, nethash distribution is good and IMO 100kH/s is longterm viable

Rotonen
2018-09-11 22:55:46
the nice thing is, that memory is only 10x quicker than NVMe

Rotonen
nvme?
what is that
2018-09-11 22:55:56
modern m.2 ssd

Rotonen
oh
right
more gbs more allocation
more speed*?
2018-09-11 22:56:34
and a quick one 512GB .. 1TB costs 100 .. 300 usd, vs. thousands in ram

Rotonen
or has to do with with
2018-09-11 22:56:44
and they take very little electricity

Rotonen
nvms speed
2018-09-11 22:56:51
well, like with ram, nand channels

Rotonen
2018-09-11 22:56:58
but as a good guideline, read IOPS / 6 -> hashrate

Rotonen
2018-09-11 22:57:29
so 500k read IOPS -> ~85kH/s + whatever fscache / preallocate gives you on top -> ~150kH/s .. 200kH/s

Rotonen
how many blocks are per day btw?
it varies?
2018-09-11 22:57:43
i recommend the 1TB sandisk for a reason

Rotonen
2018-09-11 22:58:03
well, it's random, but the network is trying to adjust the diff so it hits 144 per day

Rotonen
2018-09-11 22:58:11
aka one every 10min

Rotonen
so i assume there are big drops
every day
2018-09-11 22:58:46
was funny when block 5000 was way behind schedule and people had a panic about a potential bug :smile:

Rotonen
2018-09-11 22:59:06
yeah, some people, especially early on, were gaming the adjustment algorithm by pumping the diff

Rotonen
2018-09-11 22:59:17
that gets harder by nethash distribution

Rotonen
2018-09-11 22:59:20
having smaller nvme ssd´s add using raid 0 would be a cheaper option? You could add more if the field grows

mjay
2018-09-11 22:59:31
raid 1, until the field grows

Rotonen
2018-09-11 22:59:55
the issue is the smaller SSDs have less NAND channels

Rotonen
2018-09-11 23:00:06
currently 1TB drives seem to be the sweet spot

Rotonen
2018-09-11 23:00:20
and one should get a low power one, as they have thermal throttling issues under sustained loads

Rotonen
2018-09-11 23:00:55
I was looking at the 960 EVO 250 GB

mjay
2018-09-11 23:01:07
330k IOPS read

mjay
2018-09-11 23:01:09
that's a 10W device, i'm recommending 0,5W devices

Rotonen
2018-09-11 23:01:18
oh understood

mjay
@Rotonen can you have more than 1 nvm
2018-09-11 23:01:31
look at the 1TB sandisk 3d extreme

Rotonen
does it makes sense?
in the same machine
or 1 nvm per machine?
2018-09-11 23:02:04
depends on your motherboard block diagram, but there are also pci-e risers, but you'll still eventually hit some limit as data passes through ram, and the bandwidth saturates eventually

Rotonen
2018-09-11 23:02:34
the pci-e risers are very various in quality so cannot recommend to get into those yet before someone bites the bullet and figures out a good one

Rotonen
2018-09-11 23:02:48
but if one is so inclined to do, there's one from asus which could actually work

Rotonen
2018-09-11 23:03:06
https://www.asus.com/Motherboard-Accessory/HYPER-M-2-X16-CARD/ ASUS Global: HYPER M.2 X16 CARD | Motherboard Accessory | ASUS Global

Rotonen
Random Read/Write Speed
Up to 150,000/110,000 IOPS
2018-09-11 23:03:42
mining motherboards do not work well for that, but any multi-gpu SLI motherboard would (if one wants more than one of those)

Rotonen
is tihs any good?
2018-09-11 23:04:01
150 000 / 6 -> 25kH/s

Rotonen
2018-09-11 23:04:20
really that simple to ballpark, just divide the read iops by six

Rotonen
2018-09-11 23:04:28
SanDisk Extreme PRO 500 GB, costs half of the 1TB version, 410k IOPS

mjay
2018-09-11 23:04:40
looks like a better option to me

mjay
2018-09-11 23:05:07
dunno, 300usd is not *that* bad, and i'm not expecting anyone to go beyond 500k per m.2

Rotonen
2018-09-11 23:05:40
but sure, definitely in the correct ballpark

Rotonen
2018-09-11 23:05:42
150usd for the smaller one, 2 of those will be faster I guess

mjay
2018-09-11 23:05:56
correct

Rotonen
2018-09-11 23:06:23
and it'll take 2 new fields to force raid 1 -> raid 0, and 3 new fields to obsolete

Rotonen
2018-09-11 23:06:50
3 new fields to force me to buy another 2 :wink:

mjay
2018-09-11 23:07:04
striping reads has a management overhead, dunno if that's relevant, though, never done that on such a request density

Rotonen
350,000 IOPS
this one is the best
i think
2018-09-11 23:07:32
if the total size aligns well with the total devices size, and you have various device sizes, JBOD works well too

Rotonen
ok its not
90 eur
2018-09-11 23:08:01
500 000 to 550 000 IOPS is good, but beyond that, the power consumption should be low as well, so it does not cook itself

Rotonen
where are the big ones?
in iops?
2018-09-11 23:08:53
i use this for scouting if new interesting drives show up on the market, lessen any of the criteria to widen the search
https://geizhals.eu/?cat=hdssd&sort=r&xf=2235_500~2384_0.5~4832_3 Preisvergleich und Bewertungen für Solid State Drives (SSD) mit Schnittstelle: M.2 (PCIe), IOPS 4K lesen: ab 500k IOPS, Leistungsaufnahme Betrieb: bis 0.5W

Rotonen
2018-09-11 23:09:26
even though it's a german speaking service, the tech lingo is mostly english, so one should be able to manage mucking around with the filters

Rotonen
2018-09-11 23:09:57
above 500k there are only expensive pcie-drives, not worth it

mjay
2018-09-11 23:10:01
if you want to go beyond 500k, you need to get u.2 or the intel ruler format, but then the price is about 5k .. 10k per drive

Rotonen
can you link me your wizard again?
2018-09-11 23:10:18
sure https://snowplough.kekku.li/wizard/

Rotonen
2018-09-11 23:11:12
what i'm speculating there, is that a completely fanless low power system is somewhat optimal for snowblossom

Rotonen
2018-09-11 23:11:25
but i've not figured out how to metrify suitability of motherboards yet

Rotonen
i predict snowblossom more than 100
$
tbh
2018-09-11 23:12:18
Thats a little too much I think

mjay
its not
when snowfield 2
kicks in
actually the next one *
2018-09-11 23:12:53
the next one is 8

mjay
y
thats why i corrected
is anyone selling?
at how much?
2018-09-11 23:14:39
Not at these prices

mjay
2018-09-11 23:14:48
I doubt anyone is actually selling now

mjay
2018-09-11 23:15:06
I wait for the exchange, makes it a lot easier for me

mjay
bch exchange is cool idea
2018-09-11 23:16:14
qtrade is going to add it

mjay
can u link me
2018-09-11 23:17:33
they are still working on it, its not listed yet

mjay
any new coins like snowblossom and amoveo
?
in early stages?
2018-09-11 23:19:25
none that I know of

mjay
lets say i pay 500 $ sv for snow
and i get 10 coins per day
1 snow
has to be more than 1 $ for sure
2018-09-11 23:20:56
what hashrate did you calculate with?

mjay
none
just wondering
how much $ month
do you think its worth right now
to rent servers
and get y hashrate @Rotonen
2018-09-11 23:21:57
no renting servers is not worth it

mjay
1.6 mh
how much would you pay
for it?
month?
2018-09-11 23:22:22
Currently you get 1,05 snow per day for 100kh/s

mjay
1 snow 100kh?
no way
2018-09-11 23:23:06
50 coins / (2^38,6 / 100kh/s / 3600 / 24)

mjay
ok i get it
but how much would you pay
for 1.6 mh
month
2018-09-11 23:23:42
~510 snow

mjay
2018-09-11 23:23:51
you're asking us to speculate on future valuations, we're probably not the crowd for that

Rotonen
@Rotonen dont get me wrong im just seeing things as a "miner" now
2018-09-11 23:24:31
i throw idle capacity into mining snow since i genuinely enjoy the PoW as a technical challenge

Rotonen
2018-09-11 23:24:32
Heh. I am working on a distributed content network that leverages snowblossom

Fireduck
2018-09-11 23:24:49
renting a big-ass server is just gambling right now

mjay
better gamble on snow
:d
2018-09-11 23:25:16
have a look at used DL580 servers

mjay
2018-09-11 23:25:38
Ram is cheap for them, they have plenty of cores, and they heat up your house!

mjay
2018-09-11 23:26:04
4xE7-4870 => 2.6mh/s

mjay
2018-09-11 23:26:37
they are also rather loud

Rotonen
how much 1 of those should cost?
2018-09-11 23:27:08
depends on the deal you find, that is enterprise gear, a couple of generations old

Rotonen
2018-09-11 23:27:10
with ram below 1k usd if you are patient

mjay
2018-09-11 23:27:28
You need generation 7 btw

mjay
2018-09-11 23:28:23
And yes, they are loud :wink:

mjay
2018-09-11 23:29:03
keep in mind that you are not as flexible if the snowfield grows

mjay
2018-09-11 23:29:29
funny how the perf per watt is about the same as getting a modern small low power thing with about the same money, but of course they're ~10x faster (which is also funny how well that ram being 10x nvme holds through many examples)

Rotonen
2018-09-11 23:29:58
yeah, the absolute performance will probably drop to the same as the low power nvme miners once the next field hits?

Rotonen
2018-09-11 23:30:19
but you can just resell old servers, as you bought them, but meh, effort

Rotonen
400 $ for 1.6 mh is it any good?
x)
2018-09-11 23:30:48
thats .. per month?

mjay
y
2018-09-11 23:30:57
as a single purchase, yes it is

Rotonen
what do you mean by as a single purchase?
its per monthg
2018-09-11 23:31:29
you can get 800kH/s 256GB machines from the hetzner auctions for about 100 / month -> half the price per hash

Rotonen
2018-09-11 23:31:45
https://www.ebay.de/itm/HP-ProLiant-DL580-G7-Server-19-Zoll-gebraucht-4x-Xeon-E7-4850-10-Core-256GB-RAM/401475362520?hash=item5d79cbded8:g:udQAAOSwRE5aWOqy HP ProLiant DL580 G7 Server 19 Zoll gebraucht 4x Xeon E7-4850 10-Core 256GB RAM | Computer, Tablets & Netzwerk, Firmennetzwerke & Server, Server, Clients & Terminals | eBay!

mjay
2018-09-11 23:31:59
They are sold in large quantities, 256GB ram already

mjay
but still wouldnt be 1 cpu only better?
than 4x
2018-09-11 23:32:39
only memory channel counts matter

Rotonen
these have 4?
2018-09-11 23:33:03
16

Rotonen
2018-09-11 23:33:08
four per socket

Rotonen
6276 have 4 per socket ?
or 2 ?
2018-09-11 23:33:42
2

Rotonen
but they have 2 controllers
2018-09-11 23:33:52
or, well, depends how you count sockets

Rotonen
2018-09-11 23:33:54
i count numa nodes

Rotonen
2018-09-11 23:34:08
the dual package thing keeps throwing off my count on those

Rotonen
2018-09-11 23:34:12
yeah, they have 4 per sockets as well

Rotonen
so how much per 1 cpu?
can you estimate?
2018-09-11 23:34:41
depends on what you count as a cpu :stuck_out_tongue:

Rotonen
2018-09-11 23:35:01
the only ram mining estimate i give is 200kH/s per memory channel

Rotonen
2018-09-11 23:35:31
but realistically across a wider range of hardware that's 150 .. 250, but 200 is a reasonable baseline

Rotonen
this guy is running 4x
i will have less 2-x
can you tell by the image how many threads is he running?
68?
2018-09-11 23:37:43
no idea

Rotonen
ok 64
you can see
2018-09-11 23:38:09
start with hyperthread count, see what you get, try -50% and + 50% to see if either makes a dent, if so, explore in that direction

Rotonen
2018-09-11 23:38:32
well, htop is saying 711 threads, but dunno how many of them are active

Rotonen
i also tried the new javas
2018-09-11 23:38:54
the parallel GC for old objects can have an impact on some architectures, so try that out (there's a pointer on the wiki somewhere)

Rotonen
i tried that on vps
i will share my results tomorrow
when server is ready
2018-09-11 23:42:52
thank you, always interested in more information

Rotonen
but reserving memory
is any better
i know you answered this
but i dont remember
-XXaggressive:memory
what about this one?
@Rotonen maybe you can answer me what is "garbadge" lol
2018-09-11 23:53:54
Planning for a several million node distributed hash table is fun

Fireduck
2018-09-11 23:58:33
What is this hash table for?

mjay