allocation.txt

allocation in volunteer computing

users = job submitters
volunteers 

user types
	continuous
		always have N jobs pending
	sporadic
		submit N jobs, wait for finish, sleep M
	intermediate: batches are large

goals
	allocate resources to users in a way that performs well for all users types

client behavior
	given a set of projects, each with resource use abilities and resource share,
	allocate resources in a way that
	- maximizes throughput
	- given that, minimizes difference to resource shares
	?? what is the algorithm to do this?

===============
The proposed scheduling policy

linear model
	each user has
		balance (FLOPS)
		balance increase rate
		balance limit
			continuous users: not relevent since balance never gets high
			sporadic users: balance goes up to limit between batches.
				determines how long batches get high priority

	allocation
		indefinite: nonzero rate for indefinite period
		definite: nonzero rate for a given period

	FLOPs are measured (estimated) by
		runtime * peak FLOPS * device availability

	The sum of rates over all users should be slightly less than the actual FLOPS

	Adjust balances so lowest is always zero?

	balance constantly increases at given rate, up to limit
	when a resource becomes available (AM RPC)
		preferentially assign it to projects with high balance
		decrement by est. FLOPS over next day

===============
Other (dumber) policies

round-robin
round-robin, weighted by rate
===============
Analyzing the policy

characterize the computing resource
	don't want to model queueing in clients
	for each host
		for each resource type (CPU, GPU)
			throughput
			mean and var of added delay
		failure rate (lost jobs)?

simulator

figures of merit
	RMS completion time of sporadic batches
	continuous users gets "fair share"

questions

graphs
	balance as a function of time
		several cont users w/ different rates

	work done per project as a fn of time
	completion of batches (show retries)