How Much Memory Do You Need in 2024 to Run 1 Million Concurrent Tasks? hez2010.github.io/async-runtim... - ThreadSky

davidfowl.com • 101 days ago

How Much Memory Do You Need in 2024 to Run 1 Million Concurrent Tasks?

https://hez2010.github.io/async-runtimes-benchmarks-2024/

Comments

fakekraid.bsky.social•101 days ago

...doesn't this depend quite a bit on how generous your definition of "concurrent" is?

tfox.bsky.social•101 days ago

Feels like something fishy in the difference between thread sleep and async sleep

philip91.bsky.social•101 days ago

The task should do something before it goes to sleep or do async stuff. Then the memory footprint would be totally different.

philip91.bsky.social•101 days ago

And the task should do something random or the compiler could optimize the code and it would not show real world performance.

oliha.dev•101 days ago

NodeJS is the beige Corolla here... Just works in an acceptable range.

Go surprised me..

ryan.helegrod.dev•101 days ago

Can I ask, how would you achieve similar results in F# to the C# example, if that's possible?

I've tried a couple ways using both Task.Delay() and Async.Sleep()

The latter method consumed 11 GB!

ryan.helegrod.dev•101 days ago

Oh nevermind, I was running 10 million tasks, haha. I'm probably doing something wrong with the Async version.

The latter implementation consumes only 200 MB, which is much more in-line with my expectations.

forrestthewoods.bsky.social•101 days ago

Fun post.

> a high number of concurrent tasks can consume a significant amount of memory

Is ~400 megabytes for a million tasks significant? Ok 2.5gigabytes for Go is a bit much.

But like if you have a million tasks is that amount of memory usage genuinely prohibitive?

rubendelcampo.es•101 days ago

Also, most of the infrastructure nowadays scale memory and CPU at the same time. You can’t get a 2-core 256GB instance.

The CPU to memory ratio will make this memory consumption totally irrelevant in a real world scenario.

Interesting benchmark anyway!

m-zajac.bsky.social•101 days ago

Go always allocates 2kb of memory for a goroutine stack, because usually you are not just sleeping! Results would be totally different if the task would actually do something.

amoerie.bsky.social•101 days ago

Is that really a fair comparison if Java is starting virtual threads and C# is just adding Task.Delay instances to a list of tasks?

davidfowl.com•101 days ago

Yep

amoerie.bsky.social•101 days ago

Don't get me wrong I'm squarely in camp C# but the samples seemed structurally unequal. I guess Task.WhenAll has to do a lot of heavy lifting and that makes the comparison fair. Thanks for confirming!

bendiksens.net•101 days ago

Nice! Remember reading the old post last year.

charlieau.bsky.social•101 days ago

Impressive results from C# 👏

alexkidd71.bsky.social•101 days ago

It depends on the workload I guess.

gregsdennis.bsky.social•101 days ago

It depends on the concurrent tasks.

simplycolorful.bsky.social•101 days ago

Why does the year matter? Are tasks easier in 2024?

jabaker.bsky.social•101 days ago

Compared to 10-20 years ago?

Hell yeah, programmatic parallelism has become so much more accessible to software devs where the barrier of entry is so much easier today.

jabaker.bsky.social•101 days ago

Do you remember the days of manually writing code spinning up threads, writing your own thread pools, and dealing with synchronization primitives (granted you still sometimes with anything with shared mutability)?

jabaker.bsky.social•101 days ago

It's almost braindead easy to write code that handles concurrent tasks nowadays.

picalily.bsky.social•101 days ago

Most days I just try to figure out how many tabs I can have open 😂

ikarius.com•101 days ago

A bit surprised by Go.
I’ve been using a lot of Asyncio in Python and found it was no match to Go routines.
Weird but backed by numbers …

kashifkhan.org•101 days ago

The python drop in memory between py311 and py313 is really nice. I wonder if using a generator of tasks and passing that in to gather would further help 🤔