by Liem Nguyen, Corporate Communications Manager — January 28, 2010
I’ve been eating crow from my last post on the hapless Vikings, so it’s taking me a little bit to respond to recent blogs by Hu Yoshida and Martin Glassborow. Hu’s focuses on different aspects of tiering whereas Martin’s frames tiering within the context of disaster recovery and asks vendors to clarify how they approach risk management – both topics of much interest and debate here at Compellent. I thought I’d offer a few thoughts on performance and data protection with respect to automated tiering and how it all relates to storage virtualization, thin provisioning, replays (snapshots), boot from SAN and remote replication.
First, storage virtualization
From our point of view, storage virtualization means separating physical disk from logical volumes and virtualizing all the drives (SSD, FC, SAS, SATA) in a system into a single pool of storage. The performance of all the drives can be shared and the system spreads read/writes across all the disks regardless of type or speed. The way Compellent virtualizes storage, we don’t specify RAID ranks or RAID groups in the way other vendors do—you know who you are =), so the issue of RAID rank loss that Martin talks about doesn’t affect a Compellent customer. Loss of a RAID rank is virtually impossible. The array can dynamically move data across different storage tiers at a very granular level—Compellent page sizes are 512KB, 2MB and 4MB—enabling applications and volumes to leverage tiers between disk types and RAID levels and spindle speeds. Also, with RAID 6 we protect against two drive failures in same RAID set (more on that later).
Additional performance considerations for automated tiered storage
Hu covered a lot of the different implementations of tiered storage and their benefits. However, I think he missed a few. Namely:
- Automated tiering inside the volume – Dynamic data movement between tiers certainly cuts storage costs but if the array is tiering data only at the volume level then the efficiencies of automated tiered storage are not being maximized. Better utilization, bigger cost savings and more fluid movement result if the array manages and tiers data inside the volume – at the block level, which is what Compellent does. In a Compellent architecture, written data associated with any volume or application can tier by frequency of use so that the active blocks stay on Tier 1 and inactive blocks cascade down to Tier 2 or 3 storage. An example is e-mail. The Exchange volume doesn’t have to be associated with a specific tier of storage, which means blocks of data representing old PDF or Powerpoint attachments can move to Tier 3 (along with read-only replays of the e-mail volume) while emails about a sales deal in progress stays on Tier 1. Same volume, different tiers, no admin intervention.
- Automated tiering within each drive – Let’s take the idea of tiering from within a volume and apply it to the physical tracks on a hard drive. Writing data on the outer edges of a drive helps speed up overall performance. It’s physics really. As data is read or written, the drive head moves across the disk platters from the outer to the inner tracks and back as the disk spins. So if the arm doesn’t have to move very far, the data gets read and written faster than if the arm has to move farther in. I’m grossly simplifying but you get the idea. Some vendors place entire volumes in the outer edge to improve performance. But they’re sacrificing disk utilization and efficiency. Compellent’s Fast Track technology by itself moves the active blocks stored in every drive to the outer tracks for performance (the “fast tracks,” get it?). Inactive data gets written to the inner tracks. When you combine Fast Track with automated tiering, this block-level analysis and movement of active/inactive data happens across all the drives across all the tiers.
- Automated tiering based on spindle speed – Take SAS drives. Tiering between15K and 7200 rpm SAS, especially if you do it within the same enclosure, means small enterprises can think more seriously about automated tiered storage at lower cost. Likewise, if you started out with shelves of 10K FC drives you could just add 15K FC as Tier 1 and get more use out of the relatively slower FC drives. Or you could add SSD and SAS or SATA tiers, whichever makes sense. The point is if you’re using different speeds of the same drive type you should be able to tier between them.
- Automated tiering between RAID levels – Data movement at a block level enables tiering between RAID 10, RAID 5 or RAID 6. A lot of people don’t implement RAID 6 because of the performance overhead with writing double parity for RAID 6. In Compellent’s tiered storage, active data is written to RAID10 first and then the inactive data is dynamically moved to RAID 6. There’s a better balance of performance and fault tolerance, and the RAID6 write penalty is essentially eliminated.
In my next post I’ll cover the data protection and risk management aspects of automated tiered storage.