[Costless] - 6~11 page 대본

6 page

 

 

So this paper proposes two things :

 

we formulate the problem of optimizing the price and execution time of serverless applications. 

 

this paper proposed 2 models -> one is price model and time execution model it all includes running on cloud or devices.

 

Then with these 2 models, we represent possible function fusion and placement in a structure we call Cost Graph and formulate the problem as a Constrained Shortest Path.

 

----

 

7 page

 

 

To make a price model, this paper analyzed factors that affect the price of the lambda function. 1. Number of State Transition 2. Running or edge or cloud 3. Memory allocated to each function. 

 

-----

 

8 page

 

Lambda price is based on memory allocation, execution time, the price per 1GB memory. 

 

If we consider Lambda Function Price without state transition price. It’s $35.

 

----

 

9 page

 

 

If we consider state transition then transition cost is about $150 -> which is more costly than sum of price for executing each function. 

 

-----

10 page

 

So that’s where fusion comes in. 

 

To reduce transition costs, we can fuse the first and second functions. But It’s not effective. There is a memory discrepancy between memory. fused function’s memory allocation is set to higher memory of the previous non-fused functions. Given the latency also increased, the fused function is more costly than the non-fused one.

----

 

11 page

 

The next fusion scenario is to fuse parallel functions with their parents. If they are fused with their parents it will cause two parallel functions to run sequentially and latency of the entire workflow increases. 

 

These show why it’s important to decide which functions to fuse because it affects price and latency.