MJay

[Costless] - Revised one with transcript - p 16~20 본문

카테고리 없음

[Costless] - Revised one with transcript - p 16~20

MJSon 2019. 10. 14. 10:50

16 page

 

 

We can also fuse parallel functions with their parents

If we fuse AddtoFaceIndex, left function with its parent, CheckFaceIdDuplicate, it will cause two parallel functions to run sequentially and latency of the entire workflow increases. This suggests that proper fusion is necessary for optimizing the cost and latency.

17 page

If we want to save the cost, we can instead run the functions on edge devices because it is only charge per-device no matter how many functions we can execute. 

 

When we run function on edge device, we save intermediate data in simple storage service called Amazon S3. 

 

The problem is that it’s very slow. So when we have to run compute-intensive application, it’s better to put only a subset of the function on the edge to keep the latency with certain bounds. 

 

 

18 page

If we want to run functions on cloud from edge. There is a also transimission time to send the intermediate data stored in S3.

So we also have to think about placing the functions that reduce the transimission time on the edge device. 

 

When making the proposed model, it considers both computation and transimission time and choose the best placements, meaning which one should go to cloud or edge regarding the both price and latency. 

 

 

19 & 20 page 

 

 




Memory allocation also affects the price.

 

We set memory and AWS allocates CPU proportional to the memory configuration. However, one function ran for 4 seconds with 128MB memory, it doesn’t run for 2 seconds with 256MB Memory. Because we don’t know if the code does fully utilize CPU. 

 

So tuning the memory configuration is a difficult problem and it is important for the price and latency.

 

For memory allocation, we can think of different memory configurations as different placements of the function between edge and cloud resources.

 

Intuitively, placing a function on edge device is like placing function on VM with 128MB or VM with 512MB. The only effect is that execution time changes. 

 

So the proposed algorithm not only explores placing functions on edge devices but we can also explore placing them on 128MB cloud or 256MB cloud VM. 

 

For simplicity, this paper focuses on one cloud configuration and edge device configuration.