There is little debate anymore about the performance value of using solid state disk (SSD) in servers. They can improve boot times, provide excellent swap memory areas and accelerate the response time of many applications. But there is confusion over which type of SSD should be used in these servers. There are too many choices to be made and the enterprise is looking for a single solid state drive for all their high performance storage needs.

Right now when the enterprise selects a flash storage system there are two basic choices to be made: How should the SSD be implemented in the server and what type of flash should the SSD use.

How Should SSDs Be Implemented In Servers?

The two most common methods for implementing SSD in a server are via a PCIe card loaded up with flash memory or as an SSD, a flash storage device that looks and acts like a hard disk drive. The PCIe approach was initially very popular because of the performance it could provide. The CPU would have native access to the flash memory via the PCIe bus architecture.

Many PCIe card based flash devices require a driver in order to work; they don't use the traditional storage protocol stack. While this in theory may provide another performance advantage it is questionable how many workloads would be able to see a performance gain.

The downside to PCIe flash memory cards is that they are limited in their use case. First, there are the physical limitations that servers can only accept a limited, if any, number of full size PCIe slots, which most PCIe flash memory cards require.

Second it can be difficult to design internal High Availability (HA) with PCIe flash memory since many only support one card per server or require multiple loads of the driver software per card. Finally the native access that they enjoy via the driver causes a compatibility problem and vendors are forced to develop a driver for each operating system and/or hypervisor.

The drive form factor alternative was slow to gain acceptance in the enterprise because of concerns about the performance capabilities of the storage interfaces they connect to. Most of those issues have been resolved. Thanks to upgrades to SAS and even SATA storage interfaces and drive form factor, flash memory device should be able to provide very acceptable performance.

If acceptable performance can be obtained then there is a lot to like about an SSD standard form factor configuration. First it acts like a drive. Meaning that every operating system or hypervisor available should work with it without the need for a special driver.

Second it is relatively easy to provide HA with an SSD configuration, since to the operating system they are after all just fast drives, they can be mirrored or RAIDED as long as the operating system or file system support it. Most servers have three to five drive bays at a minimum and building a mirrored or RAIDed set of SSD should be relatively easy.

Which Type Of Flash

Flash memory is written to by changing the voltage level of a memory cell. To change the contents, the cell has to be erased (the equivalent of writing a zero) and then programmed with the actual data. A cell can only be programmed or written a finite number of times. The more frequent a cell is written to the sooner it wears out. Most flash vendors have added intelligence to the SSD controller, to make sure that data is written evenly across the device.

In the past when hard drives were selected, IT managers had to only examine price, performance and capacity. Flash technology has added a fourth point of examination: Endurance.

There are three basic types of flash technologies available on the market today: single level cell (SLC) which writes one bit per cell, multi-level cell (MLC) which writes two bits per cell and triple level cell (TLC) which write three bits per cell.

The more bits per cell that can be written, the more capacity that can be supported for the same basic cost. The problem is that more bits per cell impacts how fast the flash will wear out. So while the cost of MLC is less than SLC, it may have to be replaced more often.

This has lead data centers to have to buy different classes of flash technology for different purposes. This obviously adds to the cost by requiring higher priced SLC for write heavy environments. It also adds management complexity since wear level rates need to monitored for a variety of different devices.

Designing MLC for The Enterprise

The enterprise needs a viable server side SSD solution that can scale to meet its needs. Having multiple types of SSDs from multiple suppliers simply won't scale. Buying only SLC based flash is too expensive. The answer is to raise the reliability and endurance of MLC based SSD so that it meets the enterprise server demand in a single package.

Vendors like SMART Storage Systems are taking steps to make MLC viable to be the single SSD type for the enterprise. First they are outfitting MLC SSDs for performance by supporting the latest, highest performing SAS and SATA interface types to close the performance gap with PCIe based technology.

They are also addressing the endurance concerns by implementing a proprietary set of advanced flash management techniques called Guardian Technology to extract up to 15x more endurance from the physical flash cells. Additionally, their products leverage a technique called overprovisioning, which allocates excess capacity to the SSD to spread write traffic across all flash cells. This slows the overall rate of wear because each individual flash cell is written to less frequently.

What makes SMART Storage Systems unique in their approach is that the level of overprovisioning can be tuned via software. For enterprise environments that are going to have very heavy write workloads the level can be set very high. In essence a capacity sacrifice is made to get better endurance. For workloads that are more read heavy the level of overprovisioning can be tuned down so that more capacity is available and the cost per GB improves. The value to the enterprise is that one SSD can be selected for the enterprise as a whole, but then customized based on specific workloads.

This customization is key for the enterprise to be able to select a single SSD for qualification across multiple platforms. This reduces qualification cost and simplifies inventory management as enterprise OEM’s can stock one SKU and configure products based upon customer demand.

Conclusion

For server side flash to experience wide spread adoption it has to be cost effective, perform to expectations and have an acceptable life expectancy. Today the data center is overwhelmed with choices. The ability to provide a high performance drive with increased endurance and the flexibility to be tuned to specific workloads is an important step in standardization and broad data center adoption.

SMART Storage Systems is a client of Storage Switzerland

Previous Entry: “VM Aware Vs. ZFS Storage”