ioscale: The Holy Grail for Hyperscale The New World of Hyperscale Hyperscale describes new cloud computing deployments where hundreds or thousands of distributed servers support millions of remote, often mobile, users who expect fast, consistent response times. Driven largely by a need to support mobile and social media users, Hyperscale represents the fastest growing segment of the server infrastructure market with a 40% Compound Annual Growth Rate (CAGR) expected over the next three years. 1 Datacenter designers are racing to develop new infrastructure strategies that meet Hyperscale needs for flexible scale-out, linear budgeting, minimal power footprint, and manageability at large scale. Figure 1. Hyperscale drives new requirements. Hyperscale Demands Efficient I/O Unlike centralized scale-up enterprise datacenters, Hyperscale datacenters are characterized by distributed scaleout architectures that leverage cost-effective, high-volume servers to maximize compute efficiency. Similarly, frequent use of open-source applications eliminates expensive software licensing and allows code tuning to a company s specific needs. What has been missing is a way to scale I/O performance linearly with server performance. Centralized, proprietary SANs don t fit the cost or scaling model for Hyperscale environments and disk I/O performance has not kept up with performance advances in CPU and memory speeds. 1 ISI Group, November 1, 2012 1
ioscale Delivers Hyperscale I/O ioscale from Fusion-io boosts I/O performance and reliability to scale memory performance directly with server processing. ioscale gives Hyperscale environments the performance headroom to handle spiking holiday traffic, content gone viral, or the slash-dotted article, with far less infrastructure than mechanical disk drives or SSD-based systems. ioscale Speeds Hyperscale Application Response Times Disk-based systems, whether mechanical or solid state, place unnecessary controllers, embedded processors, and storage protocols between applications and the storage. This adds latency and slows application response times, forcing Hyperscale companies to overprovision performance for the highest expected traffic spike. The ioscale cut-through architecture minimizes latency to maximize throughput and application response times. Figure 2. ioscale provides direct access to flash devices for faster I/O performance. 2
A second benefit of the ioscale direct approach is that it can distribute the I/O workload across many more channels of flash than SSDs so each transaction processes faster. This parallel processing gives application access to more flash capacity and bandwidth than commodity SSDs. Figure 3. ioscale wide-stripes I/O so applications are more responsive. 3
ioscale Delivers Fast and Consistent Performance The ioscale cut-through architecture pipelines I/O to deliver more consistent performance than SSDs. By eliminating legacy software layers designed for slow disks, ioscale rebalances system resources so that CPUs and storage are not starved to deliver the following benefits for real-world environments: Fastest possible I/O for transactional applications and (OLTP) databases Faster servicing of I/O requests, which results in shallower queue depths and less locking and other databases blocking issues Lower processor overhead so servers need fewer cores and possibly less DRAM The end result is more predictable and consistent performance. This is critical for Hyperscale companies that must deliver high levels of performance to customers to meet stringent SLAs during peak hours. The following diagram illustrates how iomemory performance compares to SSD performance under a real-world transactional workload. Figure 4. SSD latencies stop application processing 4
ioscale Delivers More Performance Density ioscale offers more flash capacity than commodity SSD products. Higher capacity reduces server footprint and increases transaction performance per rack. ioscale delivers from 410GB to 3.2TB of low-latency, highperformance flash memory per PCI Express slot. This high capacity form factor makes each server node far more powerful, which translates into less hardware, more efficient power and cooling, and reduced maintenance costs. Figure 5. One ioscale 3.2TB delivers four times the capacity of the largest SSD 5
ioscale Minimizes Infrastructure For Better Reliability By eliminating multiple SSD devices, ioscale also improves system reliability. An SSD system requires a RAID controller to protect against device failures, and is itself a single point of failure. Figure 6. Multiple SSDs with legacy RAID controllers introduce failure points. 6
ioscale Introduces Cell-level Protection for Higher Reliability at Lower Cost ioscale Adaptive Flashback applies reliability protection at the flash cell level to minimize the impact of cell errors and failures without unnecessary redundancy and failure points. Where commodity SSDs provide protection across flash devices, Adaptive Flashback self-heals flash errors and failures with a single device. Figure 7. Adaptive Flashback protects data without adding failure points, increasing server uptime. Adaptive Flashback is an example of the benefits derived from a new architecture that is purpose built for flash devices. SSD controllers are based on disk architectures and, as a result, cannot provide this low level protection within each flash device. With ioscale, companies can achieve the same level of reliability on fewer servers, with fewer failure points and less complexity. 7
iomemory SDK Enables Flash-Optimized Hyperscale Applications Hyperscale companies have large-scale installations that can benefit from tuning storage performance to ideally match application requirements. The iomemory SDK uniquely offers an extended set of primitives for Hyperscale developers to optimize ioscale flash access. Figure 8. Atomic Writes doubled MySQL performance for Percona Percona, a leading MySQL support and solutions company, recently posted an Atomic Writes extension to the XtraDB storage engine using the Atomic Writes API. Because the MySQL Atomic Writes feature can replace double-write buffering for most MySQL updates, this extension can effectively double the write endurance storage media. 8
ioscale Offers Simple, Comprehensive Flash Management ioscale implements advanced wear-leveling technology that delivers endurance measured in Petabytes Written (PBW), and lasts longer than most components in the datacenter. The iomemory technology has proven itself in the world s most demanding environments. Fusion iosphere software provides comprehensive monitoring and management of all iomemory devices in the data center, allowing customers to see all device health and life expectancy at a glance. Figure 9. iosphere simplifies flash management for Hyperscale environments with many servers 9
ioscale Trumps the False Economy of SSDs SSDs promise more performance and lower costs. But value is about more than straight dollars per gigabyte. The SSD array approach requires more servers, more devices, and more complexity, which far outweighs any device savings. The following diagrams illustrate the efficiency and economy of ioscale-based environments. Figure 10. ioscale minimizes server and infrastructure costs. 10
Summary ioscale for Hyperscale As this paper shows, ioscale is flash that has been optimized for Hyperscale environments. ioscale allows Hyperscale customers to scale I/O performance linearly with server processing capabilities. This enables servers to deliver the fast and consistent performance required by the growing Hyperscale customer base. The availability of direct APIs allow Hyperscale developers to optimize applications deployed at very large scale for iomemory, where tuning offers large financial benefits. Figure 11. ioscale benefits propagate throughout the datacenter more effectively than SSDs. 11