Part II: Seagate 1200 12Gbs Enterprise SAS SSD StorgeIO lab review
This is the second post of a two part series, read the first post here.
Earlier this year I had the opportunity to test drive some Seagate 1200 12Gbs Enterprise SAS SSD's as a follow-up to some earlier activity trying their Enterprise TurboBoost Drives. Disclosure: Seagate has been a StorageIO client and was also the sponsor of this white paper and associated proof-points mentioned in this post.
The Server Storage I/O Blender Effect Bottleneck
The earlier proof-points focused on SSD as a target or storage device. In the following proof-points, the Seagate Enterprise 1200 SSD is used as a shared read cache (write-through). Using a write-through cache enables a given amount of SSD to give a performance benefit to other local and networked storage devices.
Non-virtualized servers with dedicated storage and I/O paths.
Aggregation causes aggravation with I/O bottlenecks because of consolidation using server virtualization. The following figure shows non-virtualized servers with their own dedicated physical machine (PM) and I/O resources. When various servers are virtualized and hosted by a common host (physical machine), their various workloads compete for I/O and other resources. In addition to competing for I/O performance resources, these different servers also tend to have diverse workloads.
Virtual server storage I/O blender bottleneck (aggregation causes aggravation)
The figure above shows aggregation causing aggravation with the result being I/O bottlenecks as various applications performance needs converge and compete with each other. The aggregation and consolidation result is a blend of random, sequential, large, small, read and write characteristics. These different storage I/O characteristics are mixed up and need to be handled by the underlying I/O capabilities of the physical machine and hypervisor. As a result, a common deployment for SSD in addition to as a target device for storing data is as a cache to cut bottlenecks for traditional spinning HDD.
In the following figure a solution is shown introducing I/O caching with SSD to help mitigate or cut the effects of server consolation causing performance aggravations.
Creating a server storage I/O blender bottleneck
Addressing the VMware Server Storage I/O blender with cache
Addressing server storage I/O blender and other bottlenecks
For these proof-points, the goal was to create an I/O bottleneck resulting from multiple VMs in a virtual server environment performing application work. In this proof-point, multiple competing VMs including a SQL Server 2012 database and an Exchange server shared the same underlying storage I/O infrastructure including HDD's The 6TB (Enterprise Capacity) HDD was configured as a VMware datastore and allocated as virtual disks to the VMs. Workloads were then run concurrently to create an I/O bottleneck for both cached and non-cached results.
Server storage I/O with virtualization roof-point configuration topology
The following figure shows two sets of proof points, cached (top) and non-cached (bottom) with three workloads. The workloads consisted of concurrent Exchange and SQL Server 2012 (TPC-B and TPC-E) running on separate virtual machine (VM) all on the same physical machine host (SUT) with database transactions being driven by two separate servers. In these proof-points, the applications data were placed onto the 6TB SAS HDD to create a bottleneck, and a portion of the SSD used as a cache. Note that the Virtunet cache software allows you to use a part of a SSD device for cache with the balance used as a regular storage target should you want to do so.
If you have paid attention to the earlier proof-points, you might notice that some of the results below are not as good as those seen in the Exchange, TPC-B and TPC-E results about. The reason is simply that the earlier proof-points were run without competing workloads, and database along with log or journal files were placed on separate drives for performance. In the following proof-point as part of creating a server storage I/O blender bottleneck the Exchange, TPC-B as well as TPC-E workloads were all running concurrently with all data on the 6TB drive (something you normally would not want to do).
Solving the VMware Server Storage I/O blender with cache
The cache and non-cached mixed workloads shown above prove how an SSD based read-cache can help to reduce I/O bottlenecks. This is an example of addressing the aggravation caused by aggregation of different competing workloads that are consolidated with server virtualization. For the workloads shown above, all data (database tables and logs) were placed on VMware virtual disks created from a datastore using a single 7.2K 6TB 12Gbps SAS HDD (e.g. Seagate Enterprise Capacity).
The guest VM system disks which included paging, applications and other data files were virtual disks using a separate datastore mapped to a single 7.2K 1TB HDD. Each workload ran for eight hours with the TPC-B and TPC-E having 50 simulated users. For the TPC-B and TPC-E workloads, two separate servers were used to drive the transaction requests to the SQL Server 2012 database. For the cached tests, a Seagate Enterprise 1200 400GB 12Gbps SAS SSD was used as the backing store for the cache software (Virtunet Systems Virtucache) that was installed and configured on the VMware host.
During the cached tests, the physical HDD for the data files (e.g. 6TB HDD) and system volumes (1TB HDD) were read cache enabled. All caching was disabled for the non-cached workloads. Note that this was only a read cache, which has the side benefit of off-loading those activities enabling the HDD to focus on writes, or read-ahead. Also note that the combined TPC-E, TPC-B and Exchange databases, logs and associated files represented over 600GB of data, there was also the combined space and thus cache impact of the two system volumes and their data. This simple workload and configuration is representative of how SSD caching can complement high-capacity HDD's
Seagate 6TB 12Gbs SAS high-capacity HDD
While the star and focus of these series of proof-points is the Seagate 1200 Enterprise 12Gbs SAS SSD, the caching software (virtunet) and Enterprise TurboBoost drives also play key supporting and favorable roles. However the 6TB 12Gbs SAS high-capacity drive caught my attention from a couple of different perspectives. Certainly the space capacity was interesting along with a 12Gbs SAS interface well suited for near-line, high-capacity and dense tiered storage environments. However for a high-capacity drive its performance is what really caught my attention both in the standard exchange, TPC-B and TPC-E workloads, as well as when combined with SSD and cache software.
This opens the door for a great combination of leveraging some amount of high-performance flash-based SSD (or TurboBoost drives) combined with cache software and high-capacity drives such as the 6TB device (Seagate now has larger versions available). Something else to mention is that the 6TB HDD in addition to being available in either 12Gbs SAS, 6Gbs SAS or 6Gbs SATA also has enhanced durability with a Read Bit Error Rate of 10 ^15 (e.g. 1 second read error per 10^15 average attempts) and an AFR (annual failure rate) of 0.63% (See more speeds and feeds here). Hence if you are concerned about using large capacity HDD's and them failing, make sure you go with those that have a high Read Bit Error Rate and a low AFR which are more common with enterprise class vs. lower cost commodity or workstation drives. Note that these high-capacity enterprise HDD's are also available with Self-Encrypting Drive (SED) options.
Summary
Read more in this StorageIO Industry Trends and Perspective (ITP) white paper compliments of Seagate 1200 12Gbs SAS SSD's and visit the Seagate Enterprise 1200 12Gbs SAS SSD page here. Moving forward there is the notion that flash SSD will be everywhere. There is a difference between all data on flash SSD vs. having some amount of SSD involved in preserving, serving and protecting (storing) information.
Key themes to keep in mind include:
- Aggregation can cause aggravation which SSD can alleviate
- A relative small amount of flash SSD in the right place can go a long way
- Fast flash storage needs fast server storage I/O access hardware and software
- Locality of reference with data close to applications is a performance enabler
- Flash SSD everywhere does not mean everything has to be SSD based
- Having some amount of flash in different places is important for flash everywhere
- Different applications have various performance characteristics
- SSD as a storage device or persistent cache can speed up IOPs and bandwidth
Flash and SSD are in your future, this comes back to the questions of how much flash SSD do you need, along with where to put it, how to use it and when.
Ok, nuff said (for now).
Cheers gs