Helping data centers deliver higher performance with less hardware

In a significant advancement for data center efficiency, researchers at MIT have developed a novel system designed to enhance the performance of storage devices by addressing three primary sources of performance variability simultaneously. This innovative approach, known as Sandook, promises to deliver substantial speed improvements over traditional methods that typically focus on just one source of variability at a time. The implications of this development could reshape how data centers manage their storage resources, ultimately leading to greater efficiency and sustainability.

Understanding the Challenge of Storage Variability

Data centers often rely on multiple storage devices pooled together over a network, allowing various applications to share these resources. This pooling is intended to improve efficiency, as not all applications require the full capacity of a storage device at any given moment. However, a persistent challenge arises due to performance variability among different storage devices, particularly solid-state drives (SSDs). Variability can stem from several factors, including differences in age, wear, and the tasks being performed on each device.

The researchers identified three main types of variability that can hinder SSD performance. The first is related to the inherent differences between SSDs that may have been acquired at different times or from different manufacturers. The second type arises from the simultaneous read and write operations on the same SSD, which can slow down data retrieval as the device must erase existing data before writing new information. The third source of variability is garbage collection, a necessary process that removes outdated data to free up space but can also introduce delays at unpredictable intervals.

To effectively tackle these challenges, the MIT team developed Sandook, which translates to "box" in Urdu, symbolizing storage. This software-based system employs a two-tier architecture, consisting of a central controller that oversees the overall distribution of tasks among the storage devices and local controllers that respond quickly to real-time conditions on each individual SSD.

The Mechanics of Sandook

Sandook's innovative design allows it to adapt dynamically to shifting workloads without the need for specialized hardware. During testing, the system was evaluated on realistic tasks such as AI model training and image compression, demonstrating an impressive ability to nearly double performance compared to traditional approaches.

The global controller within Sandook optimizes task distribution across the SSD pool, while local controllers manage immediate responses to changes in device performance. For instance, to mitigate the delays caused by read-write interference, Sandook rotates the SSDs used for these operations, reducing the likelihood of simultaneous reads and writes on the same device. Additionally, by profiling the performance of each SSD, Sandook can identify when garbage collection is occurring and adjust workloads accordingly, ensuring that devices are not overwhelmed during these critical times.

The researchers conducted tests using a pool of ten SSDs and assessed Sandook's performance across four different tasks: running a database, training a machine-learning model, compressing images, and storing user data. The results were striking, with Sandook enhancing throughput by between 12 and 94 percent compared to static methods, while also improving overall SSD capacity utilization by 23 percent. Remarkably, the system enabled SSDs to achieve 95 percent of their theoretical maximum performance without requiring any specialized hardware or application-specific updates.

Why it matters

The development of Sandook represents a pivotal moment for data centers, particularly as they face increasing demands for efficiency and sustainability. As Gohar Chaudhry, the lead author of the research, notes, maximizing the longevity of existing resources is crucial in an era where technology investments are both costly and environmentally impactful. By optimizing the performance of current SSDs, Sandook allows data centers to extract more value from their existing infrastructure before considering the need for new hardware.

The potential applications of Sandook extend beyond mere performance improvements. As the researchers look to incorporate new protocols available on the latest SSDs and leverage the predictability of AI workloads, the system could further enhance the efficiency of SSD operations in various contexts. This could be particularly beneficial in production environments where workloads fluctuate significantly.

In summary, Sandook not only addresses critical performance challenges within data centers but also aligns with broader goals of sustainability and resource optimization. As the technology matures, it holds the promise of unlocking the full potential of flash storage in cloud environments, making it a noteworthy advancement in the field of data management and storage technology.