Pisces.md

Performance Isolation and Fairness for Multi-Tenant Cloud Storage

What problem is the paper solving and why is it important?

most systems today provide weak performance isolation and fairness between tenants

What was the previous state of the art?

Network Sharing: SecondNet, Oktopus, NetShare, DaVinci

Service Sharing: Parda, mClock, Argon, Stout

Amazon DynamoDB do not provide fairness, assume uniform load distributions across tenant partitions, and are not work conserving

How does the paper advanced the state of the art?

Pisces is a system for achieving datacenter-wide per-tenant performance isolation and fair- ness in shared key-value storage. Today’s approaches for multi-tenant resource allocation are based either on per-VM allocations or hard rate limits that assume uniform workloads to achieve high utilization. Pisces achieves per-tenant weighted fair shares (or minimal rates) of the aggregate resources of the shared service, even when different tenants’ partitions are co-located and when demand for different partitions is skewed, time-varying, or bottlenecked by different server resources.

How is the system designed?

Four complementary mechanisms: partition placement, weight allocation, replica selection, and weighted fair queuing. To achieve system-wide fairness, partitions are placed with respect to demand and node capacity constraints. Local weights give tenants throughput where they need it most. Replicas are selected in a weight-sensitive manner. And request queuing is enforce dominant resource fairness

What are the key insights from the design?

Novel mechanism decomposition and novel algorithms

How is the design evaluated, what are the key results?

achieves nearly ideal (0.99 Min-Max Ratio) weighted fair sharing, strong performance isolation, and robustness to skew and shifts in tenant demand.

Open problems?

NA