Helping data centers deliver higher performance with less hardware (news.mit.edu)

MIT researchers have unveiled "Sandook," a groundbreaking software solution designed to enhance the performance of storage devices in data centers by addressing three key sources of performance variability simultaneously. Traditional methods typically tackle these issues one at a time, but Sandook's two-tier architecture features a central controller that optimizes task allocation across pooled solid-state drives (SSDs) and local controllers that manage data rerouting in real-time to overcome device congestion. This innovative approach can nearly double throughput for tasks such as AI model training and image compression, boosting SSD utilization by 23% and achieving 95% of their theoretical maximum performance without needing specialized hardware. The significance of Sandook lies in its potential to maximize the efficiency of costly and carbon-intensive storage resources, as highlighted by lead author Gohar Chaudhry. By intelligently balancing the workloads of multiple SSDs and adapting to shifting demands, Sandook not only extends the lifespan of existing devices but also improves data center performance amid varying workloads. The researchers plan to enhance the system further by integrating new SSD protocols, making this advancement crucial for optimizing flash storage technology essential for modern applications in scalable cloud environments.