Rayforce

#1

Post by tosh » Sat, Jun 06, 2026, 7:01 PM UTC

Re: Rayforce

#2

Post by noelwelsh » Tue, Jun 09, 2026, 7:09 PM UTC

I thought "morsel-driven" was AI slop, but it turns out to be in common usage in the HPC world. So I learned something from this post!

Re: Rayforce

#3

Post by arikrahman » Tue, Jun 09, 2026, 7:14 PM UTC

Any relation to the smash hit SHMUP of the same name?

Re: Rayforce

#4

Post by tosh » Tue, Jun 09, 2026, 7:29 PM UTC

noelwelsh wrote: post #2

I thought "morsel-driven" was AI slop, but it turns out to be in common usage in the HPC world. So I learned something from this post!

afaiu morsel-driven means the workload gets turned into 'smallish' chunks (morsels)

instead of having to pre-allocate upfront (e.g. 4 nodes get 1/4 each) it is more granular and dynamic

a worker that's "done" can request another morsel

pragmatic approach because nodes might not all be equally fast (cache, cpu frequency, throttling, …) and also some morsel workloads take longer than others depending on the values they contain and what kind of work needs to get done

so this approach tends to balance out nicely

I'm sure someone else can explain it better / correct me (please do!)

Re: Rayforce

#5

Post by viktorsovietov » Tue, Jun 09, 2026, 7:30 PM UTC

arikrahman wrote: post #3

Any relation to the smash hit SHMUP of the same name?

No

Re: Rayforce

#6

Post by hetoku » Tue, Jun 09, 2026, 7:44 PM UTC

noelwelsh wrote: post #2

I thought "morsel-driven" was AI slop, but it turns out to be in common usage in the HPC world. So I learned something from this post!

[flagged]

Re: Rayforce

#7

Post by hetoku » Tue, Jun 09, 2026, 7:44 PM UTC

tosh wrote: post #4

noelwelsh wrote: post #2

I thought "morsel-driven" was AI slop, but it turns out to be in common usage in the HPC world. So I learned something from this post!

afaiu morsel-driven means the workload gets turned into 'smallish' chunks (morsels) instead of having to pre-allocate upfront (e.g. 4 nodes get 1/4 each) it is more granular and dynamic a worker that's "done" can request another morsel pragmatic approach because nodes might not all be equally fast (cache, cpu frequency, throttling, …) and also some morsel workloads take longer than others depending on the values they…

Exactly!

Re: Rayforce

#8

Post by cma256 » Tue, Jun 09, 2026, 7:53 PM UTC

> Rayforce is a library you link, not a server you deploy. The C API is small enough to wrap from any language with an FFI.

I'm familiar with large-scale, commercial, client-server use cases for columnar analytics and graph traversal but what is the use case for an embedded server like this?

Re: Rayforce

#9

Post by hetoku » Tue, Jun 09, 2026, 7:59 PM UTC

cma256 wrote: post #8

> Rayforce is a library you link, not a server you deploy. The C API is small enough to wrap from any language with an FFI. I'm familiar with large-scale, commercial, client-server use cases for columnar analytics and graph traversal but what is the use case for an embedded server like this?

[dead]

Re: Rayforce

#10

Post by noelwelsh » Tue, Jun 09, 2026, 8:30 PM UTC

tosh wrote: post #4

noelwelsh wrote: post #2

I thought "morsel-driven" was AI slop, but it turns out to be in common usage in the HPC world. So I learned something from this post!

afaiu morsel-driven means the workload gets turned into 'smallish' chunks (morsels) instead of having to pre-allocate upfront (e.g. 4 nodes get 1/4 each) it is more granular and dynamic a worker that's "done" can request another morsel pragmatic approach because nodes might not all be equally fast (cache, cpu frequency, throttling, …) and also some morsel workloads take longer than others depending on the values they…

When I read up, it sounded like the same idea as work-stealing to me. Not surprising that different fields come up with the same idea under different terminology.