[Costless] - Revised one with transcript - p 1~5

Lambda provides a method state machine. The state machine specifies the orders at which lambda functions are invoked such that the output of one function is the input of the next function

 

Lambda provides a method state machine. The state machine specifies the orders at which lambda functions are invoked such that the output of one function is the input of the next function

 

 

In this case, there is also an additional cost for each transition from one to another. If we want to remove this additional cost, we can just fuse multiple functions like combining them as one function to avoid paying transition costs. However, this is not ideal if functions have different memory configurations. For example, If A function has 2GB memory 1-second duration and B function has 0.5 memory and 5-second duration, If we fuse them then, the combined function will be executed for 6 seconds with memory set as 6 seconds. This isn’t cost-effective because the B lambda function could just have been executed for 5 seconds with just 0.5GB memory. This related-problem is called function fusion problem 

Another problem they address is function placement. Instead of the lambda function, we can also use  IoT devices. AWS has a service computing ecosystem called AWS Greengrass. It’s software that extends cloud capabilities to local devices. 

 

 

Its advantage is that when functions are executed, it’s charge per IoT device rather than per function. It doesn’t matter how many functions are running on the device, the cost is fixed. ($0.16 ~ $0.22). If we leverage AWS Greengrass, we can save the cost without dramatically increasing the latency.

 

So what this paper proposes as motivation are two: 1. Function Fusion Problem and Function Placement.