nouvelles de l'entreprise Sandook delivers pooled SSD performance-boosting management

MIT and Tufts University researchers have developed a performance-enhancing management scheme for pooled SSDs that mitigates performance slowdowns caused by block erase-and-write operations and garbage collection.

Named Sandook, this scheme separates short-term and long-term performance-limiting characteristics of SSDs using a two-layer control structure fed with telemetry from agent software running on storage servers.

SSD performance can fluctuate significantly based on the mix of read and write requests. Since writes require a block-level erase and rewrite cycle, they are far slower than reads, and heavy write traffic can degrade overall read performance. Performance can also vary across SDS setups from different vendors, and even among SSDs from the same manufacturer and batch. When an SSD controller reclaims deleted cells within a block, it must copy valid data out, erase the block, and return it to the free pool—a process known as garbage collection. This operation is typically managed independently by each SSD controller and can lead to sudden, significant performance drops.

The researchers envision a compute cluster consisting of compute nodes and storage servers. A central controller operates within the compute cluster, alongside client software on compute servers and agent software on storage servers. The storage servers use standard, off-the-shelf SSDs, each equipped with its own native controller.

The Sandook Controller maintains a registry of all SSDs, having profiled each one to establish performance baselines such as IOPS capabilities. It assigns read or write modes to individual drives and receives real-time performance updates every 200ms from Sandook agents on storage servers. Using this data, it calculates read and write weights adjusted to the cluster’s global I/O demand and shares these scheduling decisions with Sandook client software on compute servers.

For reads, Sandook leverages existing block replication—already used for fault tolerance—to flexibly route read requests across replicas on different SSDs. For writes, it uses a log-structured design that allows writes to be directed to any SSD regardless of logical block placement. This high degree of flexibility ensures scheduling policies can be applied without restrictions.

Sandook client software exposes a standard block device interface to applications, transparently routing I/O requests to the most appropriate SSDs based on the controller’s scheduling rules. It also receives real-time SSD status from storage-side agents and can de-prioritize an SSD during garbage collection. Read and write requests are then redirected to other drives, preventing excessive tail latency in the overall storage system.

Sandook agents (1) process read and write commands from clients and forward them to the target SSDs; (2) provide hardware-independent monitoring, including periodic profiling and real-time status signals, to support the controller’s scheduling; and (3) relay SSD congestion signals to clients so short-term events like garbage collection can be handled locally and immediately by shifting I/O to other SSDs in the server.

The memory and CPU overhead of monitoring dozens of SSDs per storage server is minimal.

The researchers evaluated Sandook under four workloads:

Overall, Sandook improves raw I/O throughput by 30–82% compared to existing systems that only address one source of performance variability, while maintaining sub-millisecond tail latency. For unmodified applications sharing an SSD pool, end-to-end performance improves by 12–94%.

Specifically, versus prior systems, it delivers 1.7× higher storage throughput, 1.12–1.94× higher application throughput, 71–88% lower latency, and 23% higher GPU utilization—with no need for custom hardware or application changes.

The Sandook paper, titled “Unleashing The Potential of Datacenter SSDs by Taming Performance Variability,” is available as a downloadable PDF. The work will be presented at the USENIX Symposium on Networked Systems Design and Implementation (NSDI 2026) in Renton, Washington, from May 4 to 6.