SSD and HDD – ‘Friend or Foe’

There is the famous line in the Disney^TM animated classic ‘Finding Nemo’, “fish are friends, not food.” SSD providers often think of HDDs as food and not friends. That was true for laptop computers. For data center storage, SSD and HDD have many opportunities to be “friends.”

Data centers are facing increasing challenges in meeting the demands of data-intensive applications, such as big data analytics, artificial intelligence, and cloud computing. These applications require high performance, scalability, reliability, and cost-efficiency from the storage infrastructure. However, traditional hard disk drives (HDDs) are struggling to keep up with the speed and capacity requirements of these workloads, while solid state drives (SSDs) are still too expensive to replace HDDs entirely. How can data centers overcome this dilemma and achieve the best of both worlds?

One solution is to leverage the complementary strengths of SSD and HDD technologies. Data caching and performance tiering are familiar concepts in enterprise storage systems. Fast, nonvolatile write buffers have long been used to aggregate writes before striping to slower storage devices. Data tiering that offers higher levels of performance at higher cost, is common in most large storage deployments. Historically, the tiers were defined by HDDs with varying RPMs and disk diameters. Now high-performance tiers use SSDs. For instance, IDC¹ reports that for OEM enterprise storage systems in CY2023, the number of bits shipped in hybrid storage arrays (HDD + SSD), exceeded that of HDD-only or all flash arrays (AFA) by almost 2x (while the growth rate in AFA year over year was the highest by far).

Scaling Performance

For storage systems, the most important performance metric is data throughput divided by capacity (often expressed as MB/s / TB). As storage devices get bigger in capacity, they need to also scale up in MB/s of bandwidth. If they do not, the overall system performance will decrease. The required performance in a storage system varies by workload and system hardware architecture. For some typical large data center workloads, the required performance (MB/s / TB) can vary from ~2.5 for large BLOB object store, to ~5.0 for big data analytics, to ~20 for GPU clusters performing AI model training.

For HDDs, the physics and mechanics of the device fix their bandwidth. One can increase the overall throughput of the HDD by limiting the data access to exceptionally large (≥ 8MB) chunks of sequential data, i.e., maximize the data transfer time versus the time spent seeking the recording head to the data location. The gains are limited to the maximum sequential bandwidth of the HDD. It typically requires write buffering on the input to aggregate the varying size host workloads into the large chunks and caching of data for host reads that are latency-sensitive or bandwidth-constrained. The alternative approach for HDDs would be to over provision capacity, creating “dark capacity”. This approach to maintaining performance levels becomes expensive in HDD and server/infrastructure cost, and data center power requirements.

SSD and HDD Synergies

While SSDs are more expensive, they provide much higher performance and are the most appropriate storage device for these caches to bridge between the host application requirements and HDD storage subsystem. However, to achieve these gains in HDD throughput, data centers need to adopt intelligent data management software and algorithms that can automatically and dynamically allocate the data across the data tiers and caches based on the data characteristics, access patterns, and business priorities. The level of technical sophistication required to optimize and tune the management software may not be available in all data centers. The simplicity of having to deploy only one tier of storage that can meet all performance needs is the primary reason behind the high growth rate in All Flash Arrays relative to both Hybrid and HDD-only storage arrays.

These caching and tiering strategies to maximize the throughput of the HDDs have worked today for all but the most demanding workloads for HDD capacities up to 20’s TB. As HDD capacities grow to 30’s or 40’s TB in the future, it will become much more difficult to meet the workload performance needs. For instance, imagine a big data analytics application that wants 5 MB/s / TB on a 40TB HDD. The HDD would have to provide 200 MB/s of bandwidth. This is a significant challenge even in advanced data centers to ensure that a 3.5” 7200 RPM HDD can consistently maintain that speed from the outermost to the innermost parts of the disk. The relative size and data management sophistication of the SSD caches in those systems will have to increase, and hence the dependence of the HDD storage systems on the SSD performance advantages will be even greater.

Balancing cost and performance

By looking at SSDs and HDDs as “friends,” data centers can achieve an optimal balance between the cost of storage while meeting the diverse and dynamic needs of data-intensive applications. That relationship will evolve over time as HDDs get larger in capacity.

The SSDs:

• Provide fast access for hot data and latency-sensitive applications,
• Can function as cache or tier to increase the bandwidth of data transfers from HDDs,
  while HDDs:
• Provide the more cost-effective storage ($/TB),
• Have superior retention attributes for cold data storage applications, such as archival or backup data.

Longer term Futures

There will be a limit to how large HDDs can grow in capacity and still provide the most cost-effective storage for most data center applications. Eventually, SSD-only storage may become cheaper than HDD-based solutions for all but the cold storage applications. I will leave that discussion for a future blog.

^1. IDC Quarterly Enterprise Storage Tracker, 2023Q4