Many organizations are already benefiting from virtualization and the ability to consolidate multiple workloads onto a single physical server. The promise of virtualization with its favorable economics and cost savings from doing more with less has significantly changed how data centers are architected and managed. That is all good. But, there is also a persistent limitation that restricts the number and size of VMs that can be hosted on a single server: I/O. The lack of sufficient and economically available I/O has been holding back the full promise of virtualization. That is, until now.
Today, there is a simple, easy and cost effective way to dramatically extend and further scale VM density and the performance of your data center infrastructure. Adding Fusion’s ioMemory-based IO Accelerators from Hewlett-Packard and caching software to existing VMware environments overcomes inherent I/O constraints and makes it possible to run considerably more and larger VM hosted workloads on the same server. And this is accomplished without other changes made to VMware or system architecture.
Collaborating with Hewlett-Packard, Fusion-io recently completed testing of a pair of HP ProLiant DL980 G7 servers connecting with an HP P4800 back end storage system. The aim of the testing was to measure performance gains achieved from adding IO Accelerators and software to the system setup. The IO Accelerators attached directly to the server PCIe bus for use as a high-performance, flash-based read cache. The software was used to cache backend storage IOs with the IO Accelerators and also provide vMotion support.
The tests measured the system throughput as Transactions-Per-Second (TPS) using Quest Software Benchmark Factory running simulated, real-world, SQL workloads. This testing was unique with simulated database applications running on virtualized machines. User workloads typical in a Microsoft SQL server database environment were generated with instances of these workloads applied to individual VMs in increasing numbers to determine the performance characteristics in individual and aggregate TPS. The test compared the HP DL980 G7 server running VMware ESX 4.1 with and without HP IO Accelerators, and with and without software. The environment was set up with 16 concurrent VMs, each running a SQL workload.
2.8x Performance Increase with Fusion-io and the HP DL980
The test results showed that with caching software and just two 1.28TB HP IO Accelerators added to each server, an aggregated 7120 transactions per seconds were produced across the 16 VMs. This represented a 2.8x performance increase over the same baseline but un-cached configuration (2521 TPS).
What are HP IO Accelerators?
HP IO Accelerators are high performance, NAND Flash-based storage devices that attach directly to the server PCIe bus. They accelerate application and database performance by providing the server processors with low-latency, accelerated access to data, closing the gap between increasingly higher-performing CPUs and storage performance processing capabilities. Seen by the operating system (OS) as traditional block storage, it can be enabled as virtual swap space. This card utilizes and extracts performance from its flash memory using Fusion-io’s sophisticated software called VSL, Virtual Storage Layer. In addition to the HP DL980 G7 server, IO Accelerators are certified to work within 36 other HP ProLiant blade and rack servers. Card capacities range from 320GB to 1.28TB. Performance per card is: 67,000–238,000 IOPS, ~550 MB/s bandwidth and ~26 µS access latency.
What isCaching Software?
Fusion’s Software accelerates virtualized application performance and eliminates I/O bottlenecks by leveraging IO Accelerators as a transparent, acceleration device for virtualized workloads. caches data within VMware virtualized hosts, close to application workloads, delivering unparalleled I/O performance and low latency. Dramatic improvement in VM densities is enabled by distributing the IO Accelerator NAND flash memory and associated performance across the VMs so that there can be more and larger VMs per physical host. What used to be impossible or prohibitively expensive is now achievable: I/O intensive critical applications and databases can run inside VMs. And because no changes are required for VMware, organizations can deploy High Availability configurations taking full advantage of VMware vMotion. As VMs are vMotioned to or from hosts running , software dynamically rebalances capacity across IO Accelerators. This ensures that flash capacity is continuously and fully utilized and that each VM benefits from flash acceleration. In addition, user-defined controls allow designating performance acceleration to a particular VM, regardless of the number of VMs that may reside on the host at any given time.
Testing was accomplished with 16 VMs running SQL workloads in parallel on the HP DL980 and produced an aggregated 7120 TPS. In a comparison with the un-cached version of this same configuration (2521 TPS), this result demonstrates that adding just two 1.28TB HP IO Accelerators and software to the HP DL980 G7 server, performance can be increased by 2.8 times, with no other changes.
The results table for one of the 16 VMs is provided, below. The table lists the number of concurrent users, the number of transaction executions, the number of transactions per second (TPS) and the average transaction times.
sql-option 1 SQL VM driven by 1 VM with 2 agents
The time series graph, below, highlights the number of transactions per second for a single VM over a period of time, in seconds, comparing performance with and without cache. These results shown are with 100 virtual users for a single VM, but with 16 VMs running in the DL980, results at a system level will magnify this performance by a factor of 16.
Performance testing shows that the HP DL980 G7 configured with IO Accelerators and caching software can overcome the I/O constraints typically associated with virtual environments. This solution not only adds scale to the HP DL980 server but also dramatically increases VM densities to support almost any application, regardless of workload, in a virtual server environment. Having more and larger size VMs makes more efficient use of infrastructure, lowers costs and provides the opportunity to run databases in a virtualized environment, something that was not possible or cost prohibitive, until now.