Here at Teknophiles, we don’t believe in a once-size-fits-all approach to selecting hard drives for our lab servers. We prefer to adhere to the rule of specificity, where drives have a defined purpose and drive selection is based on several criteria that suit that purpose. In no particular order, we evaluate capacity, cost, reliability, performance and form factor when selecting a drive for a particular role.
Looking at this list of attributes, it’s easy to reach the conclusion that simply selecting the fastest drive would be a no-brainer for all applications. But fast drives come at an expense – both literal expense, as well as capacity expense. And, frankly, there are times where you just don’t need the the capacity or even the raw performance that some drives offer. One example, as detailed in our Silicon Power S60 60GB SSD Review, are server OS drives. On nearly every server we build, any serious workload is going to be performed on a dedicated array or SAN LUN, where IOPS and throughput are known quantities that are appropriately sized. As such, dedicated operating system drives typically experience low I/O and are approximately 75-80% read operation. You just won’t see much benefit by spending extra cash on a blazing fast SSD for your OS. And when you further consider that we nearly always configure our OS drives in RAID-1, even relatively “slow” SSDs will yield perfectly usable read speeds for a typical operating system. Heck, reliability even takes somewhat of a back seat when using cheaper drives in a RAID-1 configuration – by simply by keeping an extra $30 drive or two on hand as spares, you’ll still come out ahead financially, with no little down time waiting on a new drive to arrive.
Have Your Cake and Eat It Too
But what about those times where you do need speed, capacity, and reliability? It’s sorta like that old muscle car adage: cheap, fast, reliable – pick any two. The same premise generally holds true for computer components, including hard drives. Simply put, if you want fast and cheap, it likely won’t be reliable. Reliable and cheap? It’s not gonna be fast. You want all three? Unfortunately, you’re going to have to pay for it.
Or are you?
Perhaps there’s a happy medium – as long as you understand what it is you’re looking for. You see, performance is relative. There are certainly applications that require an abundance of IOPS. Others require significant write endurance. Others yet are heavily read biased. For each of these use cases, there are drives that fit the specific profile. For us, we needed some reliable, reasonably fast drives that will be 80% read-biased, but not break the bank. We’ve think we’ve found the sweet spot with the Samsung PM853T.
The Samsung PM853T
The Samsung PM853T series drives were mass produced around 2014-2016, so all the drives floating around out there are data center pulls, some with low hours, and in some cases, even New Old-Stock (NOS). Still, these are a great deal and can be had for as low as $0.10 per GB. Keep in mind that many of these drives are OEM drives that were sold bundled with servers, and thus will carry no warranty from the drive manufacturer even if they had a recent manufacture date. At this price point, however, having a cold-spare on hand is certainly achievable, and is highly recommended.
Samsung considers these drives to be a mixed workload drive with high sustained performance, which is perfect for our purposes. Note that the PM853T is an TLC SATA III 6 Gb/s drive, so like other SATA SSDs, it’s limited to a theoretical 600 MB/s. In our case, this is mostly irrelevant, however, as we’ll be using these in RAID-1/0 arrays as the disk subsystem for Hyper-V clusters. Given a minimum of 4 disks in an array (and possibly many more), this configuration can easily saturate the 2000 MB/s maximum throughput of a single 4-lane SFF-808x connector on an older SAS2 HBA like the LSI-Avago SAS 9210-8i.
Samsung offered the PM853T in 240 / 480 / 960 GB sizes, and the drive offers many features not found on Samsung’s consumer drives.
|Samsung PM853T – Specifications|
|Form factor||2.5 inches|
|Capacity||240 / 480 / 960 GB|
|Host Interface||SATA3 – 6 Gb/s|
|Encryption||AES 256-bit Hardware Encryption|
|Mean time||2.0 million hours|
|Uncorrectable bit||1 in 10^17|
|Power consumption||Active Read/Write : 2.7 Watt/3.8 Watt, Idle : 1.2 Watt|
– 240 GB : 150 TBW
– 480 GB : 300 TBW
– 960 GB : 600 TBW
|Cache power protection||Supported|
|Sequential R/W (MB/s)||Up to 530 / 420 MB/s|
|Random R/W (IOPs)||Up to 90,000 / 14,000|
|Physical dimensions||100mm x 70mm x 7mm|
Among the features, Samsung lists the following:
- Consistent high-quality performance. Delivers consistent performance under diverse workloads to meet various data center demands.
- Advanced Error-Correcting Code (ECC) engine. Corrects read failures to greatly improve the reliability of the data stored in the memory for higher error correction and endurance than the BCH code can deliver alone.
- End-to-end protection. Extends error detection to cover the entire path, from the host interface to the NAND flash memory in the SSD for superior data integrity.
- Power-loss protection. Ensures no data loss during unexpected power failures by using the power supply of tantalum capacitors to borrow enough time to store all cached data to flash memory.
- SMART technology. Anticipates failures and warns users of impending drive failure, enabling time to replace the ailing drive to avoid data loss and system failure malfunctions.
- Thermal throttling. Regulates the temperature of the hardware components automatically to protect them from overheating by managing its performance level to prevent data loss.
So how does it perform? Samsung provides an enormous amount of data in their product brief on the PM853T, but here are some highlights of tests conducted in Samsung’s data lab using a PM853T 480 GB drive against a competitor’s product. Samsung uses the following tools to generate this data: Fio 2.1.3, Jetstress, and IOMeter.
Sustained Performance Tests
In this test, Samsung pitted the PM853T against an competitor’s MLC SSD drive during an 11 hour workload. The results indicate that the Samsung drive shows much lower latency with less standard deviation (more consistency). Overall the Samsung drive also had overall higher average IOPS.
Additionally, the Samsung drive outperformed it’s competitor in both sustained random, as well as sequential read/write tests, achieving nearly 160000 IOPS at 100% random read in RAID-5 configurations, and 30000 IOPS at 100% random write in RAID-1 configurations.
In the sequential read tests throughput reached approximately 1500 MB/s in RAID-1 and over 1200 MB/s write in RAID-5 and outperforms its competitor as much as 29%, depending on RAID configuration and queue depth.
In mixed workloads it’s a similar story – the PM853T performs outperforms its competitor at all queue depths, in both non-RAID and RAID configurations, achieving more than 60,000 IOPs in RAID-1 at a RW ratio of 75:25, which is similar to typical virtual environment workloads.
In terms of average and maximum latency, the PM853T again performs admirably against a competitor.
Finally, in both virtual environments using multiple VMs, as well in as various real-world application workloads, the PM853T again outperforms its competitor across the board.
In our own much simpler tests, we used a Samsung PM853T 960 GB drive. This drive was a server pull that, as you can see, had very low hours.
We saw read/write performance very much in line with Samsung’s official claims and consistently saw over sequential reads over 550 MB/s read and sequential writes over 420 MB/s.
All this said, these drives do have certain limitations that should be at least touched upon:
- Form Factor. These drives are 2.5″, so they may not fit in your existing NAS, at least not without an adapter. Though at 7 mm z-height these will easily fit in all 2.5″ drive locations.
- SATA III. The PM853T is SATA III, not SAS or PCI-E, so if you need the raw performance of PCI-E or the expanded feature set of SAS such as multiple initiators, full duplex speeds or multipath I/O, then these drives are not for you.
- Write Speed. Being a read-biased drive, one would expect write performance to take a bit of a hit. These drives certainly do not display write speeds as fast as modern PCI-E/NVMe based drives. That said, they’re no slouch either, especially in RAID arrays. And at $0.10 per GB, you can actually afford to build an array with them.
- Endurance. Again, being a read-centric TLC SSD, the PM853T is only rated at 0.3 drive writes per day (DWPD). SLC drives can typically handle as many as 10x the number of write cycles that MLC or TLC drives can. This translates to nearly 300 GB in drive writes daily for 5 years. Unless you have some atypical use case, these drives should last a very long time in a typical 80%/20% R/W virtualization scenario.
So how exactly are we using these drives at Teknophiles? We’re currently running nearly 30 virtual machines on a single 1.8TB RAID-1/0 volume (4 x 960GB Samsung PM853T) and these drives don’t break a sweat. Even when hammering the environment with Windows Updates, mass live migrations or boot storms, these drives hold up well. The PM853T’s random IO performance and low latency makes it quite suitable to meet the demands of the mixed workloads that virtual machines place on the disk subsystem. Additionally, with numerous 853T drives currently in play (4 x 960 GB and 2 x 480 GB), we’ve not had a single failure in more than 10k hours use – these seem to be quite reliable drives. Simply put, for a home Hyper-V or ESX lab, it’s hard to imagine a better drive for the money. Factor in the quite excellent IO and throughput per watt of power consumption these drives produce, and you have a clear winner with the PM853T.