[Costless] 17 ~ 25 page 대본

17 page

Execution Time includes executing time on edge or cloud, transmission time. 

 

X1 includes a binary value indicating if the lambda function is executed on edge or cloud. 

18 page

 

With these 2 price models -> we can make a cost graph -> 

 

For example, Let’s say there is a function graph consisting of 3 functions

 

19 page

 

With 3 functions, we can make feasible solutions : 

 

There are 8 choices

 

First one -> all functions are run sequentially on the cloud without fusing 

 

The second one is all functions are running on the cloud with second and third functions fused

 

The fifth is first one running on edge and intermediate data are sent to cloud for running second and third functions on cloud. 

 

Note that output data can’t run from cloud to edge. 

 

20 page

With 8 possible solutions, we can make a graph. 

 

Each node represnet a solution for function placement and fusion, and each edge shows the price and latency ( execution time costs). 

 

21 page

Since price and latency have different measure units, we can’t just solve this as a simple Dijkstra’s shortest path algorithm. So we use the Constrained Shortes Path Problem. 

 

22page

 

Constrained Shortest Path Problem is applying Dijkstra’s shortest path algorithm on the aggregated cost which is made of price and execution cost. they are both normalized through dividing them by maximum cost and delay values. 

It’s for searching for the optimal value and determining when to stop the searching. 

 

----

 

23 page

This paper used the Wild Ryes application workflow. 

 

24 page

For application profiling, they ran a sequence of functions several times. They excluded the first run because of the warm-up issue. For Edge Device, it used Raspberry Pi

 

25 page

Evaluation showed that the accuracy of costless’s estimates compared to the billing information from AWS is high

 

When comparing the results of Costless with Brute force solution and other heuristics, It finds the best price within latency constraint. 

 

scheduling delay between the time of receiving the request and the time at which the function starts execution. Such delay is highly variant from one request to another and it causes the execution time to deviate by 100-300ms

We can conclude the proposed algorithm optimizes the price by more than 35%-57% with only 5%-15% increase in latency