search

fshahinfar1 / kashk

Generate BPF kernel offload from source code of applications
1 stars 0 forks source link

Fallback Strategy: Partial re-computation #11

Closed fshahinfar1 closed 3 months ago

fshahinfar1 commented 6 months ago

Currently, when we fall back, we share all computed variables with user space. This might require a lot of data copy. Moreover, not sharing may require a lot of re-computation. It seems that a partial re-computation strategy is suitable for these scenarios. In these cases. We share computation states until one point of the execution, but some value might be recalculated.

Implementing this strategy is not without challenge. Are we keeping snapshots of execution state? How to detect the good point for fallback?

fshahinfar1 commented 3 months ago

This question should be answered through the cost-model. We have a basic implementation. It does not consider the data movement, but it is possible to extend it for this purpose (probably with a moderate effort).