Understanding Computation Balancing in Distributed Systems: The Effect of Replicas on Performance

Q: If a machine stores many replicas, the original 1 GB data processing becomes several gigabytes; how much does the efficiency improve when data and computation coexist on the same machine?

A: This relates to the computation balancing problem in distributed systems . Data is typically partitioned into multiple tablets, each tablet having several replicas. A machine holds many tablets but only one replica of each tablet.

The balancing can be distilled into two basic scenarios:

When a single request accesses multiple tablets, the request is balanced across the replicas of those tablets, which are distributed on different machines; the large request is split into smaller sub‑requests, fully utilizing CPU and I/O parallelism across many machines, thus improving computation efficiency.

When multiple requests target the same tablet, they are balanced to different replicas on different machines, enhancing the computation efficiency of each request. Real‑world query processing combines these two scenarios.

Generally, a distributed system offers higher performance than a single‑node system, though the exact gain depends on the nature of the queries and workload.

However, there are edge cases where a single‑node may outperform a distributed setup due to network I/O overhead. Considering scalability, high availability, and average performance, distributed systems still provide significant advantages.