|By Louis Imershein||
|February 17, 2014 11:00 AM EST||
The initial wave of data efficiency features for primary storage focus on silos of information organized in terms of individual file systems. Deduplication and compression features provided by some vendors are limited by the scalability of those underlying file systems, essentially the file systems have become silos of optimized data. For example, NetApp deduplication can't scale beyond a 100 TB limit, because that's the limit in size of its WAFL file system. But ask anyone who's ever used NetApp deduplication if they've done it on a 100 TB file system, and you're likely to hear "are you crazy?" It's one thing to claim that data efficiency features can scale, quite a different one to actually use them with performance at scale.
Challenges around scalability generally center on two areas: scalability of random IO and memory overhead. Older solutions, like the one from NetApp, face the first challenge while newer flash-based storage systems are struggling with the second. I'll review both here:
The IO Challenge
Primary data-oriented storage devices handle both streaming and random throughput and therefore are sensitive to latency effects. Data efficiency requirements for primary storage must have fast hashing techniques to reduce the impact of latency. Fast hashes are non-cryptographic in nature and so require data comparison when used to do deduplication. It works like this:
- When a new chunk of data is read in it is first given a name using the hash algorithm.
- The system then checks a deduplication index to see if a chunk with that name has been seen before (note that this can consume disk IO and tremendous amounts of memory if done wrong).
- If the name has been seen we need to take extra steps. Because fast hashes are non-cryptographic, it is possible to have a name match while the data content differs. This is known in computer science as a hash-collision. To account for this, the existing copy of the chunk must be read in and compared bit-by-bit to the new. If they match, only a reference to the chunk is created. If not, then the new chunk must be written.
Essentially, this form of deduplication means trading a write of a duplicate chunk for a read. Depending on the design of the underlying block virtualization layer, duplicate chunks may be widely dispersed throughout the system. In that case, the bigger the system gets, the more expensive reads get - so processing of duplicate data becomes slower and slower as the storage system fills - this is why you won't find many 100 TB NetApp file systems with deduplication turned on. Certainly not for primary storage applications, the system would be flooded with random read requests and NetApp's deduplication process can end up taking months, years or even never complete.
A number of techniques have been used to reduce the impact of IO in other products. For example, the Hitachi NAS (HNAS) and Hitachi Unified Storage (HUS) solutions from HDS make use of hardware-acceleration to generate cryptographically secure hashes that do not require a data compare at all - this allows for linear scaling of deduplication performance on volumes up to 256 TB in size. Data is also written out before it is deduplicated to avoid introducing any latency through the hash computation process itself.
Permabit's own Albireo Virtual Data Optimizer (VDO) product, a plug-in module for Linux-based storage solutions, takes a different approach but with a similar result. VDO works inline to provide immediate data reduction. When data is written out, the VDO process intelligently lays it out in a sequential pattern, so that subsequent read compares of duplicates are more likely to be sequential as well. Both solutions do a fine job at solving the problem in real world scenarios, they just take different approaches.
The Memory Challenge
Many of today's flash array vendors are providing deduplication using similar fast hashing techniques to what I outlined above. With flash, the cost of doing random reads for read compares is a non-issue (random seeks on flash are much less expensive than for hard drive environments) so the use of the fast hash alone is enough to minimize latency. These systems (such as EMC's recently launched XtremIO product) are focused on delivering performance and the big challenge to performance at scale is available memory (DRAM). As above, after chunks are read in, they are named using a fast hashing algorithm. After that, the flash system must determine whether or not a chunk has been seen before. To get at this information as quickly as possible, flash-based storage systems have tended to use huge amounts of DRAM to cache chunk names in memory. It's not uncommon to see flash storage systems that allocate 16 GB of working cache per TB of storage. To support a 256 TB storage volume, such a system would require a TBs of DRAM. The increased hard costs in terms of more expensive (denser) DIMMS, as well as the increased cost of the server board required to support this many DIMMs combine to make this an extremely costly and unpopular proposition. Combine this with the fact that DRAM prices are not falling at the same rate as flash prices, and you can see why no vendor today makes a 256TB flash storage array with global deduplication capabilities.
The solution to the memory challenge is coming, in the form of a next generation of flash storage products that utilize Albireo indexing and Albireo VDO. Unlike the flash arrays described above, flash-optimized arrays with VDO takes advantage of advanced caching techniques to operate with 128 MB of working cache per TB of storage and deliver excellent performance. With VDO, a 256 TB system can be delivered with as little as 32 GB of RAM while delivering 1M IOPS performance. The net result is a cost effective and easily deployed data efficiency solution for flash arrays.
Deduplication Scalability by Vendor
As you can see in the table above, forward thinking vendors like HDS have done a good job at overcoming limitations in their data efficiency features and have products on the market today that can scale to meet the requirements of the large enterprise. Many other vendors are lagging behind, because of their inability to address IO and/or memory requirements, a serious downfall since data efficiency is at the core of distinguishing storage solutions, a critical end user requirement, and a ‘must have' component for 2014. Permabit's VDO product overcomes both of these limitations through the use of advanced memory-efficient caching techniques.
Sep. 28, 2016 09:15 PM EDT Reads: 341
Sep. 28, 2016 08:30 PM EDT Reads: 1,711
Sep. 28, 2016 08:30 PM EDT Reads: 437
Sep. 28, 2016 08:15 PM EDT Reads: 333
Sep. 28, 2016 07:00 PM EDT Reads: 1,626
Sep. 28, 2016 06:45 PM EDT Reads: 4,150
Sep. 28, 2016 06:30 PM EDT Reads: 3,116
Sep. 28, 2016 06:15 PM EDT Reads: 294
Sep. 28, 2016 06:00 PM EDT Reads: 1,774
Sep. 28, 2016 05:15 PM EDT Reads: 251
Sep. 28, 2016 05:15 PM EDT Reads: 255
Sep. 28, 2016 05:00 PM EDT Reads: 1,725
Sep. 28, 2016 04:30 PM EDT Reads: 2,816
Sep. 28, 2016 03:15 PM EDT Reads: 392
Sep. 28, 2016 03:15 PM EDT Reads: 1,314
Sep. 28, 2016 03:00 PM EDT Reads: 3,877
Sep. 28, 2016 02:45 PM EDT Reads: 1,746
Sep. 28, 2016 02:30 PM EDT Reads: 2,933
Sep. 28, 2016 02:30 PM EDT Reads: 3,339
Sep. 28, 2016 02:00 PM EDT Reads: 4,427