Sandook: A Breakthrough in Data Center Efficiency

By Sophia Wang;

Physics and Aerospace Associate; The Lawrenceville School, NJ

Given the rising demand for data centers, a new system developed by MIT researchers to optimize their performance has the potential to transform the tech industry. Typical methods used to increase efficiency such as acquiring advanced hardware or upgrading old infrastructure can be expensive and impractical. However, this breakthrough in software provides a way to optimize data centers at a comparatively lower cost. The program, called Sandook, addresses  current flaws in performance simultaneously, unlike current programs which offer only single-pronged solutions, opening new possibilities for growth in the field of technology. 

Solid-state drives (SSDs) play a central role in modern data centers, serving as storage devices that are reliable, low-latency, and energy-efficient. Their main tasks include managing and authorizing the reading and writing of data, as they quickly retrieve and store information. Typically, SSDs are grouped together and their capacity is shared amongst multiple applications. This boosts efficiency, as an application may not require an SSD’s full capacity at all times. However, this aggregation carries a large downside: due to the large variability of performance across different SSDs, inefficient units can impede the whole pool’s ability to function.

First of all, this variance arises from the wide range of models, wear levels, and capacity between SSDs. For instance, Samsung, Meta, and Alibaba versions have had extensive operative discrepancies. Another source of variation stems from differing read-to-write ratios required by SSDs for different tasks. Before writing, the SSD must first erase data to free up space, thus impeding the efficiency of concurrent reading or retrieval processes. Finally, sporadic “garbage collection” activities regarding the deletion of data requires multiple read, write, and erase-oriented tasks that occupy SSDs, causing performance decline.

Current efforts to address these drawbacks only tackle one form of variance at a time. Furthermore, fixes to individual issues may be incompatible with one another, as many carry contradictory clauses regulating SSD function. 

Sandook, on the other hand, uses a two-tiered framework to resolve  these issues by controlling load distribution over SSDs. It can account for  variabilities and allocate load accordingly to optimize efficiency. The first layer consists of a general scheduler that distributes load steadily with a global view, while the second layer utilizes a case-by-case approach, reacting quickly to pressing events and sudden SSD congestion. As a result, Sandook can direct load on the microsecond as well as on the long-term level.

To determine load distribution, Sandook employs an SSD agent to monitor the performance of each individual SSD by tracking its latency over incremental increases in load, thus profiling its projected performance under varying workloads. This screening occurs regularly to account for changes in function over time. Additionally, the agent uses real-time latency metrics to detect congestion – such as that caused by garbage collection – and subsequently reports this data back to Sandook. 

At the Sandook controller level, the system organizes, reads, and writes, preventing them from interfering with each other. It considers the performance levels of different SSDs and shifts their workloads accordingly using linear programming. Furthermore, when Sandook receives a congestion signal from the SSD agent, it decreases the load on that device before gradually increasing the load back to its primary working level. Hence, these elements work together to maximize efficiency.

The researchers found that applying Sandook increases application throughput, which measures the rate of successfully processed requests, by 12-94%, latency by 71-88%, and SSD utilization capacity by 23%. 

All in all, this breakthrough has the ability to transform data centers, paving the way for growth in AI. Sandook unlocks doors to more affordable and accessible technology, potentially bridging the gap between the present and the future in regards to the usage and application of AI.

References

Chaudhry, G. I., †, Bhardwaj, A., Ruan, Z., Belay, A., MIT CSAIL, & Tufts University. (n.d.). Unleashing the potential of datacenter SSDs by taming performance variability. In MIT CSAIL [Journal-article]. https://goharirfan.me/publications/sandook_nsdi_2026.pdf

Helping data centers deliver higher performance with less hardware. (2026, April 7). MIT News | Massachusetts Institute of Technology. https://news.mit.edu/2026/helping-data-centers-deliver-higher-performance-less-hardware-0407

Ibm. (2025, November 17). Solid State Drives. Think. Retrieved May 5, 2026, from https://www.ibm.com/think/topics/solid-state-drives

MIT Energy Initiative. (2026, February 27). Data center power demand. Main. https://energy.mit.edu/strategic-priorities/data-center-power-demand/

Pavlovic, D. (2024, December 17). What is an SSD? The Ultimate Guide to Solid-State Drives. HP® Tech Takes. https://www.hp.com/us-en/shop/tech-takes/what-is-an-ssd