Improve Performance and Efficiency With Hybrid and All-Flash Storage
Many z/OS clients benefit from performance and efficiency gains by implementing flash storage.
By Nick Clayton01/02/2017
The storage industry is seeing a transition from traditional HDD to SSD using NAND flash technology. Adoption of this technology for high-performance transactional systems has reached a tipping point. Many z/OS* clients benefit from performance and efficiency gains by implementing flash storage.
The IBM DS8000* family includes both the all-flash DS8888, and hybrid DS8884 and DS8886 systems; this article describes the options available and provides tips for effective planning and implementation. It also outlines the differences between z/OS and distributed systems, and how other IBM z Systems* technologies—such as zEnterprise* Data Compression (zEDC)—complement the use of flash storage.
Flash provides significant latency improvements compared with traditional HDDs. However, it’s still slower than dynamic random access memory, so read and write cache provide significant benefits. Write cache helps coalesce and avoid repeated writes from unnecessarily wearing out the flash media. The Intelligent Write Caching algorithms on the DS8000 storage developed by IBM Research consider both temporal and spatial locality to minimize wear on the flash.
Flash can also sustain significantly higher random I/O throughput than traditional HDDs. Flash-optimized hardware, such as the high-performance flash enclosure in the DS8000 family, is important to ensure RAID systems designed for HDD don’t become a bottleneck when using flash. However, in most cases, applications aren’t actually driving the maximum random I/O throughput in flash environments but are still significantly above what’s possible with traditional HDDs.
Equally important to z/OS environments is good sequential performance for batch workloads. Modern flash media are significantly superior to earlier generation technology in this aspect—especially for sequential writes. However, a flash-based system must handle the required sequential workload as well as random I/O.
The density of flash storage has increased over the last three years to the point where significant density savings exist when compared with high-performance enterprise drives. Depending on the cost of data center floor space and the associated power and cooling, this can have a major impact on the overall affordability of flash storage.
Response Time Shrinks
In addition to the speed of the underlying storage media, several factors affect a storage system’s host response time. With the implementation of flash storage, the media response time can become a smaller percentage of the overall response time. Minimizing the other components then becomes critical to making further reductions in response time.
Moving to faster SAN fabrics, channels and host adapters can help reduce the overall host response time. Another often overlooked factor is the contribution of the protocol and software stack on the host to the overall response time. High Performance FICON* for z Systems has provided significant improvements in this area, and the DS8880 includes other enhancements to reduce latency and improve throughput for specific workloads. These include IBM zHyperWrite* for DB2* logs, the DB2 Castout Accelerator and the List Prefetch Optimizer.
Overall cache hit ratios in z/OS environments tend to be significantly higher than distributed systems, however average values can hide variation from one application to another. In many cases, specific transactions or batch jobs receive a higher benefit than others. With flash, the benefit is avoiding the requirement to spend valuable time and skilled resources on application performance tuning and letting flash’s low latency handle those cache misses.
The latest generation of z Systems servers supports larger memory sizes, and middleware (e.g, DB2, IMS* and CICS*) can exploit this to provide larger amounts of caching in server memory. This reduces the cache hit ratios on the remaining I/O, meaning historical assumptions about the cache hit ratios for z/OS will likely not apply in the future.
A core value proposition of the mainframe platform is the capability to run analytics workloads directly on systems of record rather than performing extract, transform and load processing and then performing analytics on a copy of the data. This can drive additional I/O workload on the same data volume; without flash storage, this can cause performance issues. Clients who have implemented flash storage for z/OS can enable analytics such as fraud detection on the same systems and in real time while still meeting performance goals.
Compression and Deduplication
In distributed environments, many storage systems offer compression and/or deduplication capabilities to reduce the relative price of flash storage. IBM offerings for this include the A9000 and the Storwize* platform.
On z Systems, the zEDC and compression coprocessor provide application-aware compression capabilities that can be used no matter the type of storage media. Therefore, significantly reduced benefit exists in implementing compression in the storage system, as compression higher in the stack benefits the server and SAN as well as back-end storage capacity. In the case of the zEDC, this can provide superior compression ratios as well.
Continued Cost Reduction
Flash storage media costs continue to decrease 25 to 30 percent per year—faster than traditional HDD. Many workloads are also past the point where larger HDD is viable; in these cases, the cost of HDD storage may not actually decrease over time.
When comparing the costs of different media types in an enterprise storage system, examine more than just the cost per gigabyte of the media, as this can provide an unrealistic view of the differences. Larger drives spread the cost of I/O enclosures, device adapters and the rack infrastructure over a larger number of terabytes of data, which can significantly affect the relative costs.
For example, a 3.2 TB flash card (rated at 10 drive writes per day) is around 3.1x the cost per terabyte of a 300 GB 15K enterprise drive when considering only the cost of the media. When the other required infrastructure is included, it’s only 1.5x higher in terms of acquisition cost. This calculation doesn’t include costs of disk controllers, host adapters and advanced function licenses or floor space/power/cooling and maintenance costs, so the differences are lower when looking at the total cost of ownership (TCO) of the system.
Hybrid Systems and Tiering
While an all-flash system can provide incredible price-performance for costs similar to 15K enterprise drives, further cost reduction may be desired. In these situations, a hybrid system with flash cards combined with larger 10K and/or nearline drives can provide significantly lower costs—and, with Easy Tier* functionality, better overall performance than 15K drives. Configurations with 20 to 40 percent flash and larger capacity 10K drives are now attractive from both a cost and performance perspective.
Easy Tier technology provides automated storage tiering at a subvolume level, placing busier, “hot” data on faster storage and less busy, “cold” data on cheaper, higher capacity storage. With release 8.1 of the DS8880 storage, Easy Tier enhancements take into account flash costs and the higher percentages of flash storage being deployed.
It’s now possible to configure Easy Tier technology as the home tier for new allocations. This means that newly created volumes or newly allocated extents of thin provisioning will get the best performance when first used and then will later be demoted if appropriate.
Easy Tier is also more reactive when promoting data to upper tiers. Previously, the analysis for this promotion was done daily, but now it evaluates the activity on the lower tiers every five minutes. If evidence of overload exists, it immediately promotes data.
Anyone deploying new storage systems for z/OS should be looking to include flash storage in their configuration. All-flash configurations deliver improved throughput and performance with significant density and data center environmental improvements, and are attractive from a TCO perspective where predictable performance for all workloads is a key focus.
For cost-sensitive environments, hybrid systems can provide reduced cost and increased density compared with traditional enterprise drives, and will provide equivalent or better performance.
Nick Clayton is an executive IT specialist and solution architect for DS8000 development, IBM Systems.
Sponsored ContentSoftware-defined Secondary Storage Solutions Come of Age
Post a Comment
Note: Comments are moderated and will not appear until approvedcomments powered by Disqus