'Slow' is a loaded term with a lot of different definitions.

In terms of throughput, there's nothing wrong with a garbage collector, but in terms of latency (particularly 95+ percentile) you're never going to be able touch 'manually' (I'm including Rust here) managing the heap. Even Azul's 'pauseless' gc only guarantees something like 10ms max pauses. That's an eternity for a lot of use cases.

In a lot of use cases, like the drivers mentioned above, you shouldn't allocate on the heap anyway, whether manually or in a managed way. And if you don't allocate, you never need to garbage collect (since you never need to reclaim memory to make it available for new allocations). In that case, the two are even.

(The above assumes that GC is only triggered by allocations. Many GCs are like that, but others trigger periodically even if there isn't anything to do. I don't know which group OCaml is in.)

You heap allocate all the time in drivers. You just shouldn't do it from interrupt context. Source: I write plenty of drivers for a living.

And even if the GC is only triggered by allocations, they'll still pause the other contexts in order to scan their stacks for live references. So, in the case of a unikernel, you can end up with critical paths being paused by allocations outside the critical path.

> You heap allocate all the time in drivers. You just shouldn't do it from interrupt context. Source: I write plenty of drivers for a living.

Fair enough. I was thinking of low-level things like "shovel a few bytes from some hardware component into a user-provided buffer" without allocating anything, you seem to be working on more high-level drivers.

> And even if the GC is only triggered by allocations, they'll still pause the other contexts in order to scan their stacks for live references.

If the GC is only triggered on allocations when there isn't enough free heap space, then why would stack scans happen at other times?

For a driver, you have many contexts you're running. That's what the windows error "IRQL not less than or equal" blue screen means, that someone called a regular kernel function from within interrupt context. In drivers, you almost always heap allocate in your non interrupt contexts in order to dynamically handle load. This would be even more necessary in a unikernel, where you don't have a user space to defer that kind of work to.

What I'm saying is that the necessary heap allocations happening in other contexts will block progress of your non allocating critical path, irq code on a GC based unikernel because of the need to discover liveness information. There are potential schemes to fix this, but they aren't implemented by MirageOS/OCaml, or most other managed unikernel environments I've seen.

But if I understand correctly (I've never used mirage) mirage is a unikernel for ocaml applications. It compiles down to xen system calls, for monitoring and administration. Mirage isn't actually an implementation of an OS in the traditional sense.

Ocaml is of course already fast. INRIA has been using ocaml for compiler hacking for several decades. The speed comes as no surprise, but eventually rust will have inline assembly support, it's already nurturing some basic SIMD features, it is already being turned for lower level dominance.

I think ocaml could compete, but not the way it is now.

With linear types in the future, perhaps ocaml could follow rust's lead and implement a memory model around move semantics.

I'm not an expert in this area, but I do perceive a nontrivial gap here. Ocaml is not built for writing system drivers. And a good system driver is written in something that was designed to be used for doing so, imo.

Sounds like you may like https://news.ycombinator.com/item?id=16597329

Oberon and Blue Bottle drivers written in Oberon and Active Oberon respectively.

http://www.projectoberon.com

https://svn.inf.ethz.ch/svn/lecturers/a2/trunk/

Mirage TCP/IP drivers

https://github.com/mirage/mirage-tcpip