looked into openmpi again for miner tuning, and urgh, javac wrappers and all the trimmings https://www.open-mpi.org/faq/?category=java
oh god, mpi
that takes me back
interconnects, funfun
i think arktika will do for now
for quite a while :smile:
the great, I am a scientist who can barely program but I need a bunch of machines to work on this problem language
if it could fail in a bunch of strange ways, that would be a big plus
oh, the archaea of computing duct tape of academia
all the fun is exiting from those stacks as universities form joint companies to do shared data centres to do things proper
i think your 80s cluster approach to within-miner heterogeny is actually the way most things are going to work in the end
one job I had was at a place that could checkpoint and restart mpi cluster jobs
which of course the software having no idea
since it was likely some night mare code no one could change
sounds like some shared memory flush side effect hack is the way there
like a distributed ACPI hibernate
it was basically rolling partition of state through the cluster. Things on one side got suspended if they tried to talk to the other.
as long as you can intercept every IO call and suspect all the memory state, it worked nicely
suspend
Anyways, Arktika could use some discovery mechanisms to make it easier to setup
or at least notice if something is wrong
but if you write the configs right and don't screw up it works pretty well
oh, discovery mechanisms, my favourite ones were from a chinese MMO we localized for some european market, the data centre got angry at us when it blasted some 400Mbps worth of UDP for hours across the data centre just to find the game nodes
instead of doing something sane like mDNS
rockin
I'd use multicast and assume it is probably a single local network
sad how multicast is advanced futuristic still
hah, yeah
ran a student network in 2005, we had iptv multicast going then
and if someone has some multicast aware routers and spread it around, that is fine too
and hdtv since 2007
hpe, cisco and juniper compete to sponsor that network - that was not an issue... :smile:
nice
also ISPs and TV channels liked to use it as their field lab for multicast
if we can just get everyone on ipv6...
~5k users
dunno, the zeroconf stuff is scary
it seems like it would be easy to advertise your node as an ipv6 router and start collecting traffic
or, let's say zeroconf being used to software define OSI L2 is scary
actually ipv6 is probably as safe or even safer for the arktika zeroconf as multicast is, might even be likelier to work well with whatever people have in their home networks
I mostly use ipv6 on my home network
but my automatic config consists of machines putting a dns entry into route53 after they get an ipv6 address
.ch is sadder in that, i got a symmetric 10G fibre at home, but cannot get ipv6 for that
way too many local services have AAAA records, but nothing sensible answers there, so cannot do that to consumers
like cities, media outlets, the like
maybe you should just use netbios?
IPX
yeah, kali sucks, so gotta go the whole mile