I really like the manycores approach, but we haven’t seen it come to fruition — at least not on general purpose machines. I think a machine that exposes each subset of cores as a NUMA node and doesn’t try to flatten memory across the entire set of cores might be a much more workable approach. Otherwise the interconnect becomes the scaling limit quickly (all cores being able to access all memory at speed).
Erlang, at least the programming model, lends itself well to this, where each process has a local heap. If that can stay resident to a subsection of the CPU, that might lend itself better to a reasonably priced many core architecture.
> think a machine that exposes each subset of cores as a NUMA node and doesn’t try to flatten memory across the entire set of cores might be a much more workable approach. Otherwise the interconnect becomes the scaling limit quickly (all cores being able to access all memory at speed).
Epyc has a mode where it does 4 numa nodes per socket, IIRC. It seems like that should be good if your software is NUMA aware or NUMA friendly.
But most of the desktop class hardware has all the cores sharing a single memory controller anyway, so if you had separate NUMA nodes, it wouldn't reflect reality.
Reducing cross core communication (NUMA or not) is the key to getting high performance parallelism. Erlang helps because any cross process communication is explicit, so there's no hidden communication as can sometimes happen in languages with shared memory between threads. (Yes, ets is shared, but it's also explicit communication in my book)
> Erlang, at least the programming model, lends itself well to this, where each process has a local heap.
That loosely describes plenty of multithreaded workloads, perhaps even most of them. A thread that doesn't keep its memory writes "local" to itself as much as possible will run into heavy contention with other threads and performance will suffer a lot. It's usual to try and write multithreaded workloads in a way that tries to minimize the chance of contention, even though this may not involve a literal "one local heap per core".
Yes, but in Erlang, everything on every process is immutable and nothing is ever trying to write anywhere besides locally. Every variable assignment leaves the previous memory unchanged and fully accessible to anything directly referencing it.
Paraphrasing the late great Joe Armstrong, the great thing about Erlang as opposed to just about any other language is that every year the same program gets twice as fast as last year.
Manycores hasn't succeeded because frankly the programming model of essentially every other language is stuck in 1950. I, the program, am the entire and sole thing running on this computer, and must manually manage resources to match its capabilities. Hence async/await, mutable memory, race checkers, function coloring, all that nonsense. If half the effort spent straining to get the ghost PDP-11 ruling all the programming languages had been spent on cleaning up the (several) warts in the actor model and its few implementations, we'd all be driving Waymos on Jupiter by now.
I'm curious, which actor model warts are you referring to exactly?
[The obvious candidates from my point of view are (1) it's an abstract mathematical model with dispersed application/implementations, most of which introduce additional constraints (in other words, there is no central theory of the actor model implementation space), and (2) the message transport semantics are fixed: the model assumes eventual out-of-order delivery of an unbounded stream of messages. I think they should have enumerated the space of transport capabilities including ordered/unordered, reliable/unreliable within the core model. Treatment of bounded queuing in the core model would also be nice, but you can model that as an unreliable intermediate actor that drops messages or implements a backpressure handshake when the queue is full.]
I don't think either of those are particularly problematic. The actor model as implemented by Erlang is concrete and robust enough. The big problems with the actor model are, in my opinion, around (1) speed optimizations for immutable memory and message passing (currently, there's a great deal of copying and pointer chasing involved, which can be slow and is a ripe area for optimization), (2) (for Erlang) speed and QOL improvements for math and strings (Erlang historically is not about fast math or string handling, but both of those do comprise a great deal of general purpose programming), (3) (for Erlang) operational QOL misc improvements (e.g. existing distribution, ets, amnesia, failover, hot upgrade, node deployment, build process range from arcane (amnesia, hot upgrades, etc.all the way up to covered-in-terrifying-spiders (e.g. debugging queuing issues, rebar3))
There is no lineage between The Actor Model and Erlang. The creators of Erlang are on record as having never heard of the Actor Model (as developed by Hewitt, Agha and colleagues at MIT). None of the points you make (including the first one) are a part of any formal definition or elaboration of the Actor Model that I have seen, which was one of my points: there is no unified theory of the Actor Model that addresses all of the practical issues.
With respect to your point (1), you might be interested in Pony, which has been discussed here from time to time, most recently: https://news.ycombinator.com/item?id=44719413 Of course there are other actor-based systems in wide use such as Akka.
Can you explain the joe armstrong quote a bit to someone not familiar with the language?
Erlang's runtime system, the BEAM, automatically takes care of scheduling the execution of lightweight erlang processes across many cpus/cores. So a well written Erlang program can be sped up almost linearly by adding more cpus/cores. And since we are seeing more and more cores being crammed into cpus each year, what Joe meant is that by deploying your code on the latest cpu, you've doubled the performance without touching your code.
> Erlang, at least the programming model, lends itself well to this, where each process has a local heap. If that can stay resident to a subsection of the CPU, that might lend itself better to a reasonably priced many core architecture.
I tend to agree.
Where it gets -really- interesting to think about, are concepts like 'core parking' actors of a given type on specific cores; e.x. 'somebusinessprocess' actor code all happens on a specific fixed set of cores and 'account' actors run on a different fixed set of cores, versus having all the cores going back and forth between both.
Could theoretically get a benefit due to instruction cache being very consistent per core, giving benefits due to the mechanical sympathy (I think Disruptors also take advantage of this).
On the other hand, it may not be as big a benefit, in the sense that cross process writes are cross core writes and those tend to lead to their own issues...
fun to think about.
The BEAM launches a scheduler process per CPU thread in SMP mode, although I don't know if it moves Erlang processes between them.
The behavior is configurable and the default is unbound.
https://www.erlang.org/doc/apps/erts/erl_cmd.html#%2Bsbt
Who knows what will really happen, but there have been rumours of significant core-count bumps in Ryzen 6, which would edge the mainstream significantly closer to manycore.