HDD vs SSD: an easy choice?
On paper, many choices seem initially clear. A sports car seems to offer significantly more performance on paper than a run-of-the-mill vehicle. But its shortcomings become clear once it is time to take the kids to school, bring home the shopping from the supermarket, or collect furniture from the local store. And, for most of us, the choice is one or the other, not purchasing both so that the limitations of one can be offset by the other when required.
Author: Rainer W. Kaese, Senior Manager, Toshiba Electronics Europe
In the world of data storage, a similar dilemma applies. On paper, SSDs are clearly the winner, offering heaps of performance, whereas the well-known HDD alternative, although cheaper and offering more capacity per device, is the less-appealing of the two. But this paper comparison fails to consider getting behind the wheel and actually driving these two storage-media on a daily basis.
In recent years, flash-based storage, such as SSDs, have developed to provide huge capacities in ever-smaller form factors. Not being mechanical, they are rugged and robust for use in laptops and other portable devices. Without the need to seek out sectors on a spinning platter, they are also significantly more responsive than HDDs and typically provide higher throughput when compared side-by-side for input/output operations (IOPS). HDDs on the other hand have focused on mass storage, growing to sizes of 16 TB per drive and with 20 TB on the horizon.
The introduction of helium as the gas used inside the drive, along with continued improvements in spindle drive technology and the disk controller, mean that the remaining specifications, such as noise, power consumption, and performance, are continually improving too. However, it is clear that HDDs will never compete against SSDs purely on the basis of noise, performance or power consumption.
With the obvious side-by-side comparison tipped towards a preference for SSDs, the next key issue is the cost. The advantages of SSDs come at a significant price premium, currently being up to ten times the cost of HDDs per TB when comparing devices of the same class. Some applications will justify this expense, such as the boot drive of a server or other tier-0 applications where a single to a handful of drives make sense. Of course, this also leads to the consideration that a system fitted with both SSDs and HDDs could be constructed that provides the best of both worlds, offering a blend of high performance and masses of storage.
However, at scale, the reality looks different, especially when working within the confines of a defined budget. Most storage applications need to find a good balance between total capacity and performance, especially in the common application areas of web hosting, email serving, generic, virtual and cloud storage, backup and archiving, and content delivery. Using a fixed budget, Toshiba undertook an investigation to see at what point the move to HDDs provided more performance than a system based solely upon enterprise SSDs (eSSD).
A system based upon eight 1.6 TB SATA eSSDs connected to a Microsemi Adaptec SuperRAID controller in a Supermicro chassis was built. Configured as a RAID6 array to provide as much total storage as possible, the system offers a net capacity of around 10 TB but with double parity for data protection. The drives selected also offer 3DWPD (drive writes per day), the minimum recommended for active enterprise storage applications.
For the same investment, a comparable system was constructed of 24 10,500 RPM, 2.4 TB SAS HDDs. As the system was competing against high performance SSDs, these were configured in a RAID10 array. This allowed the parallel RAID0 striping to aggregate performance while the RAID1 mirroring delivered data protection, albeit with a 50% overhead for that protection. The RAID controller came from the same supplier as before, with a chassis for the larger quantity of drives in a 2U hot-swap JBOD coming from AIC. This configuration provided a total storage of around 30 TB.
Random workloads were created using the Flexible IO tester ‘FIO’, an open-source synthetic benchmarking tool for testing differing workloads on storage systems. This provided a method for comparing like with like for differing block sizes on the two systems. As was to be expected, the eSSD solution delivers higher performance and throughput at small block sizes of between 4k and 32k bytes. The HDD platform lies anywhere between 1.6 and seven times behind the eSSD platform for IOPS in this range.
However, at block sizes of 64k byte and above, the performance impact of the higher number of HDDs starts to show through. IOPS improvements of 14% were attained, rising to 86% if the block size were to increase to 512k byte.
Of course, it can be countered that real-world applications do not consist of a single block size, but even splitting the system 20/50/20/10 across 4k, 64k, 256k and 2M byte blocks respectively, the HDD server continues to provide 37% higher IOPS compared to its eSSD counterpart.
Further research also shows that, for the same budget, 60 2 TB SATA 7,200 RPM HDDs can be arranged in a RAID10 configuration and deliver even higher IOPS. Compared to the eSSD arrangement at 64k byte block size, IOPS lie 69% higher, and it even offers a 48% improvement over the server fitted with the 10,500 RPM HDD drives. All this comes with at a total capacity of around 60 TB, double the earlier HDD example and six times that of the eSSD solution.
Data centres are of course increasingly concerned about power consumption and large quantities of HDDs obviously cannot compete with the low power demands of eSSDs. However, taken on aggregate, the HDD solution does not look too bad when it is also considered that the performance over eSSD is improved and that there is significantly more available storage. Returning to the two original examples, the eSSD solution would require around 788 kWh per year against 2015 kWh per year for the HDD. Assuming €0.07/kWh and a Power Usage Effectiveness (PUE) of 1.3, the HDD solution only requires €110 more in energy costs per year, equating to less than 10€ per month extra.
Data centres are not built or upgraded every day, so the diligent manager needs to also factor in the technology improvements that are likely to come in the foreseeable future. Flash technology seems to be moving at a rapid pace and it is fair to question whether SSDs may reach a price point that makes them more attractive in that time-frame. The reality is that the industry manufactured around 800 EB (Exabyte – 800Mio TB) of HDD storage in 2018. But only around 100 EB were manufactured by SSD suppliers.
Currently, flash memory manufactures are pouring investment into ramping up capacity, but current projections indicate that $100B - $200B will be required to double the existing flash memory capacity. At current rates, that will only cover one quarter of existing consumption. With HDDs earmarked to reach 20 TB per drive in the next 12 – 18 months, it is clear that SSDs will not be achieving significant reductions in cost-per-terabyte in the foreseeable future, nor will there be a significant improvement in their cost-per-terabyte pricing.
It is clear that there are specific applications and usage cases where SSDs have a clear advantage over HDDs, and it is likely that that number will increase over time rather than decrease as the capability vs capacity vs price comparison is made. What is certain is that HDDs have a clear role to play and for a long time to come.
The continued improvement in power consumption and performance, although limited, will help the scales to fall in the HDD’s favour, while the price per terabyte and the capacity per drive will always be preferable for HDDs. What this research clearly shows is that system performance for the same outlay under conditions that, for many data centres, will remain core business, HDDs can significantly outperform SSDs at compelling total capacities for a minimal additional energy cost outlay.