Putting It All Together

Candidate and GroupAdjust work together to combine differentiated service with other system-wide performance goals. Figure 7 illustrates how they allocate surplus memory to optimize global response time for four example services. In this example, the hosted services have identical $\lambda$ and caching profiles $(\alpha, S, T)$, but their SLAs specify different response time targets. Candidate allocates the minimum memory (total 46 MB) to meet the targets, assigning more memory to services with more stringent targets. GroupAdjust then allocates surplus memory to the services that offer the best marginal improvement in overall response time. Since the services have equivalent behavior, the surplus goes to the services with the smallest allotments.

Figure 8 further illustrates the flexibility of MBRP to adapt to observed changes in load or system behavior. It shows allotments $[M, \phi ]$ for three competing services s0, s1, and s2. The system is configured to consolidate all loads on a shared server and storage unit, meet minimal response time targets when possible, and use any surplus resources to optimize global response time. The services begin with identical arrival rates $\lambda$, caching profiles $(\alpha, S, T)$, and response time targets. The experiment modifies a single parameter in each of six steps and shows its effect on the allotments; the changes through the steps are cumulative.

• In the base case, all services receive equal shares.

• Step 2 doubles $\lambda$ for s0. The system shifts memory to s0 to reflect its higher share of the request load, and increases its $\phi$ to rebalance storage utilization to $\rho_{target}$, compensating for a higher storage load.

• Step 3 triples the arrival rate for s1. The system shifts memory to s1; note that the memory shares match each service's share of the total request load. The system also rebalances storage by growing s1's share at the expense of the lightly loaded s2.
• Step 4 reduces the cache locality of s2 by reducing its $\alpha$ from .9 to .6. This forces the system to shift resources to s2 to meet its response time target, at the cost of increasing global average response time.

• Step 5 lowers the response time target for s2 by $40\%$. The system further compromises its global response time goal by shifting even more memory to s2 to meet its more stringent target. The additional memory reduces the storage load for s2, allowing the system to shift some storage resource to the more heavily loaded services.

• The final step in this experiment reduces the amount of system memory by $10\%$. Since s2 holds the minimum memory needed to meet its target, the system steals memory from s0 and s1, and rebalances storage by increasing s1's share slightly.