In the traditional storage market, high-performance and high-reliability deployments have been dominated by 10,000 and 15,000 RPM SAS hard-disk drives (HDDs). Whether it be server-attached or distributed storage networks, SAS HDDs have played a pivotal role in the growth of storage over the past 10 years. In 2015 alone, analysts predict that nearly 33 million 10,000 and 15,000 RPM HDDs were shipped. This is in addition to a nearly 136 million unit install base.
While solid-state drives (SSDs) have made heavy inroads into other storage areas, such as personal computing, SAS SSDs remain a small fraction of the enterprise market. Cost is the main hurdle that has prevented broader adoption. Even though the demand for faster, higher capacity storage is increasing, typical IT budgets are flat. That, combined with the general impression of SSDs being cost-prohibitive has stunted adoption.
With the release of the Micron S600DC series SAS SSDs, pricing is no longer the hurdle that it once was.
When designing a storage system, there are many factors that contribute to the total cost of ownership (TCO). Acquisition costs are typically the key consideration as they are visible line items in the current year’s budget and are easily tracked. Other factors can be just as important but not nearly as visible.
In this brief, we will discuss the following items that contribute considerably to TCO:
- Acquisition Costs
- System Costs
- Power and Cooling
- Physical Space
The value proposition of SAS SSDs in data centers are fundamentally changing. No longer will you have to weigh features like performance and reliability against cost-per-gigabyte. You don’t have to spend time and resources deploying a handful of SSDs to augment HDDs in complex, tiered storage solutions. Micron’s S600 series SAS SSDs deliver the best of everything: better performance at a lower cost than previous HDD deployments.
When compared to SSDs, even the fastest HDDs are painfully slow, as shown in Table 1 below. Whether you are talking about large block sequential performance or small block random performance, SSDs can be orders of magnitude faster depending on the condition. This should not be news to anyone working in IT – it is a wellestablished fact. How much faster SSDs are is key. Input operations per second, or IOPS, is a metric you will see on every SSD data sheet and is a basic measure of how many READ or WRITE transactions can be completed in a given second.
1 Based on SAS SSD data sheets.
2 Based on 3rd party reviews.
3 Based on 3rd party reviews.
4 Inner to outer diameter performance (performance data).
5 Inner to outer diameter performance (performance data).
Table 1: SSD Performance Comparison
Table 1 shows that for small transfers, as seen in database transactions, the S600 series reaches 200,000 random read and 120,000 random write IOPS. While HDD vendors do not directly quote IOPS, many online review sites put the typical 10,000 RPM HDD at ~400 IOPS and 15,000 RPM HDDs at ~500 IOPS, best case.
For large block, sequential operations, such as content delivery, the S600 series is rated at 1.9 GB/s for read operations and 850 MB/s for write operations. In contrast, HDDs can typically only sustain 250MB/s, read or write, while accessing the outer rim of the disk, and dip close to 100 MB/s at the inner rim.
One common practice to increase the speed of HDDs is to short stroke the drive. Short stroking involves limiting the capacity of the drive in order to speed-up random operations. By limiting the area that the head of the disk has to travel, the latency to access the data decreases and the number of operations performed per second increases. The tradeoff is that you are paying for more capacity than you are actively using, effectively increasing the cost per gigabyte. Unfortunately, this practice will only increase performance by 10’s of percent, which is still orders of magnitude less than an SSD.
For applications that rely on random read performance, the S600 series is up to 400 times faster than the fastest HDD on the market.
One of the challenges SSDs have faced in the past is the limited amount of drive capacity. When dealing with large datasets, limited capacities means more drives are required for a given use-case. Increasing the number of drives can cause a cascading effect of more systems, more rack space, more networking and more power and cooling. The S600 makes a much more compelling case for SSD capacities.
Table 2: Drive Capacities
For higher performance systems, users are currently limited to 600GB for a 2.5” 15,000 RPM SAS HDD. Systems using 2.5” 10,000 RPM HDDs are only slightly better off, topping out at 1.8TB. Contrast that to the S600 series that boasts a maximum capacity of 3.84TB, which is 2-4 times higher than competing HDDs in the same formfactor. Unlike HDDs, which are growing capacities very slowly, new NAND technologies are paving the way for 8
and 16TB SAS SSDs in the near future.
As previously mentioned, the largest limiting factor for SAS SSD adoption is acquisition costs. Historically, the typical cost per gigabyte comparison has skewed strongly in favor of HDDs. Only a few years ago, SAS SSDs were close to 10 times the price of competing SAS HDDs. While SSDs have always had a massive performance edge, that initial cost was difficult for IT budgets to absorb.
With the S600 series, Micron is able hit a price point that is at parity or lower than 15,000 RPM HDDs1. Let that sink in for a minute. We are now at a point where the acquisition costs for SSDs are equal or lower than [certain/some/high-end] HDDs for the first time ever.
1 Based on CDW.com pricing, February 2016.
When architecting a system, IT professionals have to look beyond simple drive-to-drive comparisons. They need to consider the system as a whole and all the associated costs that come with it.
- Pricing for 15,000 RPM HDD 600GB based on CDW.com
- Pricing for Dell R730 based on Dell.com at the time of publication
- Pricing for S610DC based on distribution pricing at the time of publication
Figure 1: HDD and SSD Comparison
Our first example is a simple server-attached storage system used for online transaction processing (OLTP). For OLTP systems, low latency and high IOPS is desirable because they dictate the number of transactions that can be processed per minute. These systems typically use 15,000 RPM HDDs to provide that performance.
In this example, we have 16 600GB 15,000 RPM HDDs in a RAID 10 configuration. RAID 10 is used because it offers excellent performance along with parity in the case of disk failure. It is preferred over RAID 5 and RAID 6 because they both have long rebuild times if a drive were to fail. The cost for that performance and parity is capacity: RAID 10 reduces the usable capacity by 50%. Each drive has also been short stroked to 450GB to squeeze every last bit of performance out of the array.
With the Micron S610DC, we are able to match the parity level (RAID 1), but also increase the usable capacity (from 3.6TB to 3.84TB). The total cost of the system was also reduced by 27% (from $14,630 to $10,670). The real value, though, comes when performance is added to the equation. The S610DC system provides over three times the random write performance and 40 times the random read performance. In other words, customers that are focused purely on performance and not total capacity would need to spend $43,890 to match the $10,670 S610DC system.
In the next example, we scale our S610DC system up to 8 drives to keep the RAID 10 parity. For our 15,000 RPM HDD system, we are using 300GB drives in RAID 10 without short stroking.
- Pricing for 15K HDD 600GB based on CDW.com
- Pricing for Dell R730 based on Dell.com
- Pricing for S610DC based on distribution pricing
Figure 2: HDD and SSD Comparison
To match the capacity of the S610DC system, you would need 96 300GB 15,000 RPM HDDs in 4 external 24-drive JBOD enclosures, and even then, you would have almost a terabyte less capacity. Even though we changed a few of the parameters, the total cost of the S610DC system is still 27% less than the 15,000 RPM HDD system.
Power and Cooling
When evaluating power usage, there are three areas that need to be addressed: Active Power, Idle Power and System Power.
Comparing active power draw between HDDs and SSDs is not always straightforward. HDDs draw power consistently, regardless of the workload, as there is always a base power level required to spin the disk and move the actuator arm. SSDs draw different amounts of power based on the workload and that difference can be substantial. In general, sequential workloads draw more power than random workloads and write heavy workloads draw more power than read heavy workloads. So, while there may not be much difference in the power draw between an HDD and SSD during write operations, read operations on an SSD can consume less than 50% of the power when compared to a HDD.
Figure 3: Active Power Comparison—Source: Published SSD Datasheets on Micron.com and Product Data
When a SSD drive is not actively working, it will typically enter an idle state. According to product datasheets an idle SSD can consume upwards of 2W less than a competing 15,000 RPM SAS HDD. Since SAS SSDs are so much faster than HDDs, another important metric is the time it takes to complete a given task. For example, it takes 40 seconds to write 10GB of data to disk sequentially using a 15,000 RPM HDD (based on the sequential write specifications in Table 1). For the S650DC, that task would only take 11.76 seconds. That means that the S650DC could be at idle for over 28 seconds while the 15K SAS HDD is still working. The longer any drive can spend at idle, the less power it will consume. In this example, the 15K HDD requires 64% more power than the S650DC. The power savings may seem trivial for a single drive over a short time span, but when you extrapolate that out to thousands of drives working over many years, the power savings can be substantial.
The gap in power requirements becomes a canyon you examine power at the system level. In Example 2, 8 S610DC SSDs can replace 96 15,000 SAS SSDs with 4 external JBOD enclosures. When accounting for just the storage subsystem and assuming maximum power draw for all drives, the HDD system requires nearly 15 times the power of the SSD system (1075W vs 72W).
Figure 4: System Power Draw Comparison —Source: Published SSD Datasheets on Micron.com and Product Data
In the enterprise market, customers are becoming much more cognizant about power requirement implications. Not only does power draw increase your operating expenses, the resulting cooling requirements compound the issue.
Another hidden cost of large HDD deployments is the physical space they occupy. Example 2showshow a single server with 8 S610DC SSDs (2U of space) can replace 4 external JBOD enclosures with 9615,000 RPM HDDs (10U of space). While the reduction of 8U’s worth of space in a single rack may not require any additional floor space, the scale-out implications can be enormous.
Figure 5: Physical Space Diagram
Imagine if you had 20 SSD systems that would fit into a single 42U rack with 2U to spare. Using HDDs, that same system would require 200U of rack space, requiring 4 additional racks. For data centers in dense urban areas, that extra space can come at a premium cost.
Drive reliability is another area where SSDs hold a clear advantage at both the drive and system level. Typically, drive reliability can be viewed in two different ways: Mean time to failure (MTTF) or annualized failure rate (AFR). Both of these metrics rely on the same base data provided by the manufacturer and only differ by the representative units. MTTF is the measure of average service hours between failures among the total install base. AFR is the average failure rate per year among the total install base.
For a 15,000 RPM SAS HDD, the MTTF is typically 2 million hours or a 0.44% AFR. Micron’s S600 Series SAS SSDs are rated at 2.5 million hours MTTF or 0.35% AFR (Sources: Published SSD Datasheets on Micron.com and Seagate Product Data). The lack of moving parts (disk, actuator arm, etc.) is just one of the reasons that SSDs have superior reliability.
At a system level, as shown in Example 2, the vast number of HDDs required only exacerbates the issue. Even if individual drive reliability was equal, the HDD system requires 12 times the number of drives which reduces its reliability significantly. When you add on external JBOD enclosures and power supplies, the problem only gets worse.
One of the biggest hurdles SSDs have in enterprise environments is endurance. Due to physical properties of the underlying NAND, SSDs have a finite amount of data that they can write throughout their lifetime. It is easy to focus on the fact that SSD wear differently than HDDs. But, when you look at the total amount of data that is capable of being written, SSDs still hold a distinct advantage.
For SSDs, endurance is typically rated in the number of drive writes per day (DWPD). This is a measure of how many times the full capacity of a drive can be written each day throughout its warranty period. For example, the Micron S610DC 3840GB can sustain 7008TB of writes during its lifetime (3840GB x 365 days x 5 years x 1 DWPD).
Unfortunately, HDDs are not typically specified in the same way, but the following example provides a better comparison. A 15,000 RPM HDD, regardless of capacity, can sustain 500 random 4KB write IOPS. That means that over five years, the maximum amount of data that it can write is 323TB, or 22 less than the S610DC.
Even though the 15,000 RPM HDD might be able to handle more writes than an equivalent SSD, the fact that the media is so much slower limits the amount of data you can write in its lifetime. Table 2 shows how this gap widens as we compare to higher endurance SSDs.
Table 3: Endurance Comparison—Source: Published SSD Datasheets and Capable TBW and DWPD
Calculations Based on IOPS Capable for 10,000 to 15,000 RPM HDDs
If we look at it in DWPD, the S610DC is rated at 1 DWPD while the equivalent for a 15K HDD is only 0.29 DWPD, or a 3X improvement. The reason why there is a difference between TBW and DWPD is that DWPD takes into account capacity while the calculation for HDDs assumes a constant write speed across capacities. The “capable” DWPD for HDDs will typically go up as capacity decreases.
There are a few near-line, 7200RPM HDDs that do rate their endurance, but it is important to read the fine print. These HDDs are rated at up to 550TB of data per year or 2750TB over five years. This may seem like a lot but that figure is based on data transferred to or from the drive. What this means is that even in a read-centric workload, you are building towards that endurance number. Comparing this to SSDs, where only writes count towards their endurance specification. For these HDDs, if you go beyond this number, you risk increasing the annualized failure rate for the drive.
Transitioning to new storage technologies is always a difficult proposition. With the S600 Series SSDs, Micron is making that transition as easy and as smooth as possible. Compared to 15,000 RPM SAS HDDs, Micron S600 Series SSDs provide superior performance, capacity, acquisition costs, system costs, power and cooling, physical space, reliability and endurance.