Are SSDs inching past being used just for cache to speed up IOPS, or could they soon be an option for tier 1 storage?
With the introduction of IBM i 6.1, a number of features were introduced to the operating system to support solid-state drives (SSDs). The new "solid" drives were faster, quieter, and more efficient than conventional spinning discs. Alas, they also were more expensive. The solution that IBM and others suggested to maximize speed but contain costs was to integrate SSDs with hard-disk drives (HDDs) and allocate the most frequently used data to the faster drives and relegate the rest to the conventional disk drives.
The task of automatically separating "hot" from cold data is easier said than done, and companies have been founded based on their algorithms that allow systems to put more frequently sought data and applications onto SSDs while leaving cold, or less frequently used, data on HDDs. Nevertheless, IBM has found a way to handle this in IBM i and has enhanced the feature further with the release of IBM i 7.1.
But is anyone really using solid-state drives on Power Systems computers, and, if so, are they realizing any benefits? While real concerns still exist about the price of SSDs, there seems to be no question about their ability to speed up I/O operations. Partial payback of some of their cost can be realized through savings in other areas, such as energy and consolidation of storage hardware, including controllers. While some have questioned their reliability, IBM insists that SSDs—at least the enterprise class drives it sells for use with Power Systems—are just as reliable as more conventional HDDs. Yes, the SSD will wear out after about five years, but five years is a long time. Since a new one can be initialized in a matter of minutes, as opposed to hours for an HDD, popping in a replacement should not be a big deal.
Solid-state drives have been around for awhile, but they weren't considered economical or reliable enough for enterprise storage until about three years ago. The SSD uses solid-state memory to store persistent data that it offers up to the user similarly to a block I/O HDD. The data is stored in microchips, and SSDs have no moving parts. They are quiet and less likely than electromechanical drives—with their moveable read/write heads—to be damaged falling off an airport conveyor belt. This makes them ideal for use in laptops. Best of all, they have far lower access times and less latency than spinning discs.
I spoke with Mark Olson, an IBM Power Systems World Wide Product Manager, about SSDs on Power and also got a briefing from Don Chouinard, Nexsan Director of Product Marketing, on Nexsan's latest series of NST 5000 unified storage appliances employing their FASTier Cache solution that relies on SSDs. Both men said that SSDs were one of the hottest trends in storage today, a trend fueled by descending prices, increased reliability, and businesses' need for speed in a climate of expanding data.
Chris Smith: "Don, what have the SSDs done for the storage industry? They're fairly new and weren 't being used to any great extent two years ago, were they?"
Don Chouinard: "No, and you know why? Wear spots; they wear out. When you go to use SSD as primary storage, the cost is through the roof. If you want 24 200GB SLT SSD, you best have your banker at your side when you buy that system. And you are going to have to discard it in five years. So people didn't really want to go that way, so now what the big storage vendors have done is said, 'We're going to put your actual data onto rotating media. It's very cost-effective, and we're going to use RAID. The drives will kick out here and there, but RAID will control those little blips, and you're not going to have to throw it out in five years; it will run indefinitely. We're then going to accelerate your rotating media using these SSDs. Now if an SSD kicks out, we just stop using it for caching. We mark it as bad, and work around it, and pull it out and replace it—it's only cache. For writes, you just make sure your writes are always saved, and for reads, you just go around an SSD that may be breaking down.' The cost on SSDs is coming down, and the reliability and density are going up. But you have to be careful because the really dense SSDs are MLC, or multiple bits per cell, and they are nowhere near as reliable as a single bit per cell. Twenty years from now, there may not be any rotating media, but we're going to have to solve the fact that these SSDs are just dead in five years; they're like a light bulb. They go out and have to be replaced. People are not accepting that, so they have found their place as cache in front of rotating media."
Nexsan's FASTier Cache involves using a write journal next to the server that records transactions later drained and applied to the underlying rotating media. It also utilizes dynamic random access memory (DRAM)-based SSDs on the serial attached SCSI (SAS) bus and contains a second copy of each controller's write journal. Finally, 100 and 200GB SLC Flash-based SSDs on the SAS bus cache in-demand blocks to accelerate all rotating drives using opportunistic read-aheads and write-behinds. The result is an appliance that can triple the performance of serial advanced technology attachment (SATA) and SAS drives giving SATA drives the same random I/O performance as a 15,000 rpm SAS drive. Deployed largely in Windows environments, the NST5000 series also serves AIX and Linux customers.
IBM External Storage
IBM has its own high-end enterprise storage system enhanced with SSD drives. The IBM XIV Storage System Gen3 has been designed to support mixed workloads running on virtual server environments, offers high performance for various I/O profiles, is simple to provision, manage, and tune, is scalable and even capable of serving in a service-centric, private cloud infrastructure. Distribution of data across the disk drives is automatic and requires no manual tuning. It features various data-efficient technologies and by June will include transparent solid-state drive caching that will leverage SSDs as a caching layer between each XIV module's RAM and the underlying SAS-connected disk drives.
For midsize businesses, IBM has the Storwize V7000 Unified Disk System, which has been designed to be easy to use and fast to deploy. It offers virtual storage with enhanced efficiency and flexibility through built-in solid-state drive optimization and thin-provisioning technologies. Its advanced functions also enable non-disruptive migration of data from existing storage. Its virtualization capabilities allow for the incorporation of existing as well as new storage systems.
The V7000 Easy Tier technology allows the user to deploy SSDs with confidence by automatically and dynamically moving only the appropriate data to the SSDs in the system based on ongoing performance monitoring.
IBM Internal Storage
For other customers, who want to take advantage of their internal storage on their Power Systems machines, IBM recently introduced a totally new PCIe2 1.8 GB Cache RAID SAS Adapter (#5913). It is a large-cache PCIe SAS adapter that provides high-performance capabilities for large quantities of SSDs or HDDS, according to IBM.
The company says it is the most advanced SAS controller ever offered, allowing users to reduce the number of PCI slots required and thus help minimize the overall I/O configuration. To support its higher-speed interface and high-density ports, new cables also were introduced.
Mark Olson: "We do have a range of options you can pick from when configuring servers. You might think of them in three different classifications.There are SSDs that plug into PCIe slots, and then we have SSDs that go into the SAS space, kind of like disk drives except we put SSDs there [SAS-bay-based SSDs], and then we have SAN SSDs. The new fast adapter we brought out in the fourth quarter—announced in October —has a design with SSDs in mind. It's the first SAS adapter we've ever done that way, and there is some industry-leading technology there. The adapter really hums. You can attach a bunch of SSDs to it and get great performance. Employing a smaller number of PCI slots, you can have a lot more SSDs. Your footprint then shrinks, your energy improves, and the overall cost per configuration improves. It is available for both POWER6 and POWER7 machines, and it is a really sweet option."
Continues Olson: "The adapter is a double-wide card, so you plug it into two adjacent PCIe slots and then there are four bays, or slots, in the card itself where you can slide in SSD modules. The modules are the size of little fat credit cards and are the 1.8 inch drives. They just slide in; there are four of them. We can get up to 708GB of space with four cards, unless you're doing RAID 5, which could take away from the effective space the customer has. We really like the adapter approach. We had all the infrastructure in place to use a SAS controller, which we designed at IBM. We already had the drivers and the SAS controllers. Though we couldn't drive huge numbers of SSDs, driving just four of them works pretty darn well. In that adapter, we also provide a huge amount of function that outside vendors don't provide. If you look at some of the drivers that they offer, they take cycles from the processors that do the work."
Summing It Up
Olson is more bullish on using SSDs for tier 1 storage and challenges the notion that all SSD drives fail immediately upon five years' service. "We design it to last five years with incredible write usage for 24x7x365," says Olson. "No one uses SSD technology in a real-life environment like that. But even if they did, it's a disk drive. If it starts to weaken, and if we start to see that it is running out of writes and that the drive is weakening, then if you're under IBM maintenance, it gets replaced."
Olson also challenged the idea that single bit per cell, or SLC, technology is "far more reliable" than MLC, or multiple bits per cell. "The only drives we sell now are MLC technology," says Olson. "People must understand…that there's a difference between generic MLC technology that you can go down and buy for your PC…and the enterprise-class MLC technology that we use in Power Systems," he says. The notion that SSD drives are any less reliable than HDD drives is, at this point, not reasoned thinking emanating from measurable criteria but instead is a myth based on anecdotal experience of people using lower-grade, inexpensive drives. "The vast majority of the vendors in the industry as a whole are working on MLC—not SLC," says Olson.
As far as future trends, Olson says he sees "improvements in price/performance" that will continue. As the technology "gets more mature," the cost per gigabyte and cost per I/OP are both going to improve. Packaging will improve with flash memory using smaller circuitry that will continue to drive up the amount of storage in a given physical footprint. "It will be very similar to how disk drives have come down in terms of physical space for an ever-increasing amount of storage capacity. SSD drives will have the same level of reliability, often with increased speed through the inherent Flash technology.
"From the controller perspective, we'll continue to enhance those as well, leveraging faster controllers, more optimized controllers. I think you're going to see advances in both areas within the next 18 to 24 months," says Olson.
SSDs clearly are gaining headway, whether or not for tier 1 storage, then certainly for cache-type storage in front of rotating HDDs. If reliability is no longer an issue for enterprise customers, then only price is in the way of much broader adoption of these impressively fast drives. If predictions by Olson and others are accurate, and prices continue to drop, expect to see SSD drives as a regular part of nearly every Power Systems server.