.png){kind=icon}
.png){kind=icon}
Return
The Cache Oracle: Preemptive Request Dispatching for I/O-Asymmetric Serverless Databases
.png){kind=icon}
PRE
AI
DOI：10.1002/cpe.70470.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In hyperscale, multi-tenant serverless databases, the substantial performance disparity between cache hits (microseconds, CPU-bound) and cache misses (milliseconds, I/O-bound) leads to severe interference among tenants. Traditional reactive resource management models, which respond only after an expensive cache miss occurs, fail to prevent I/O-intensive noisy neighbors from degrading the latency experienced by well-behaved tenants. This paper proposes a novel preemptive isolation approach: the Predictive Cache-Aware Dispatcher (PCAD). Embedded within the database proxy layer, PCAD leverages a real-time machine learning model-the Cache Oracle-to predict the cache state of each request prior to dispatch. This foresight enables two key optimizations: (1) preemptive routing of likely cache misses to a dedicated, concurrency-limited I/O worker pool, thereby protecting the CPU-optimized fast path from contention, and (2) speculative prefetching to hide disk access latency for predicted misses. Using production traces from a large-scale multi-tenant cloud database, we demonstrate that the Cache Oracle achieves over 91% F1-score in cache miss prediction. Simulations show that PCAD effectively eliminates cross-tenant interference caused by I/O bursts, sustaining stable sub-millisecond P99 latency, while speculative prefetching reduces perceived cache miss latency by over 90%. This preemptive predict and prevent approach marks a significant advancement toward more robust and high-performance autonomous database systems.
Keywords:
machine learning for systems
multi-tenancy
performance isolation
serverless databases
Journal
C
IF:
1.5
Papers:
473
Citations:
0