Technical marketing white paper NEW HP 3PAR StoreServ Flash Storage Not Just For Mission Critical Application Acceleration Anymore
Introduction Enterprise and mid-tier flash solid-state drive (SSD) storage have quickly emerged as the go-to storage standard for mission-critical performance hungry applications. Whether it is online transaction processing (OLTP); virtual desktop interface (VDI); Enterprise resource planning (ERP); virtual machines (VMs); server clustering; customer resource management (CRM); high frequency or derivatives trading; film & video post production; rendering; reservoir modeling; seismic processing; protein decoding; genome decoding; flow analysis; climate modeling; simulations of all kinds; or just about every application that is latency or IOPS sensitive, Flash-SSD storage has proven to be a tremendous boon to its performance. Users are happier. IT administrators are happier. So what s the downside? Naysayers contend that smart utilization of caching and fast HDDs can match most of the flash SSD performance gains. They argue that flash SSD storage is neither as reliable nor as durable as hard disk drive (HDD) storage. And that SSD storage cannot match HDD density or capacities. But mostly they protest that Enterprise and mid-tier storage utilizing flash SSDs cost much too much to justify their performance gains especially for the vast majority of data center applications that are not IOPS hungry or performance sensitive. How accurate are these common objections? Do they hold any water? This paper details, analyzes, deconstructs, and critiques each of the objections while establishing how they do not apply to HP 3PAR StoreServ Flash Storage. It will establish how HP 3PAR StoreServ Flash Storage reliably meets the needs of both mission critical IOPS hungry and latency sensitive workloads as well as all of the other primary data center applications at a cost point lower than HDD equivalent storage. It will close with real use cases that demonstrate how HP 3PAR StoreServ Flash Storage leveraging new cost-effective high capacity flash SSDs from SanDisk can make a profound difference. 2
Introduction... 2 Performance Hungry Applications... 4 Flash SSD Storage Mythology Inhibiting Adoption... 5 1. Reliability and Endurance... 5 2. Density... 7 3. Total Cost of Acquisition (TCA) and Total Cost of Ownership (TCO)... 8 What All This Means In The Real World... 9 For More Information... 10 3
Performance Hungry Applications Applications that are performance hungry generally need low latency storage that translates into high IOs per second (IOPS). This is especially true for mission-critical applications that rely on databases such as OLTP, VDI, ERP, CRM, VMs, and more. Meeting that relentless rising demand for IOPS from traditional HDD based storage systems has become progressively more challenging, complicated, costly, and tenuous. HDDs are electro-mechanical devices with spinning drive platters and a drive head that must fly (move) over the disk platter until it is in the right location. IOPS per GB is affected primarily by the speed of the drive platters; next by the density or capacity of the drive; size of each IO operation (.5K, 2K, 4K, 8K, 64K, etc.), and finally whether the IOPS are sequential or random. HDDs don t spin any faster than 15,000-RPM because higher speeds radically increase vibration because of little things like sonic booms (going faster than the speed of sound) inside the drive. Increased vibrations increases data errors and drive failures. HDD spin is not getting any faster so raw IOPS out of any given drive is limited. And while capacities are in fact growing, it has the counterintuitive effect of reducing the IOPS per GB (HDD non-increasing IOPS/increasing Capacity = < IOPS/GB. Finally, most of the IOPS intensive mission-critical applications are random IOPS and will vary in size between applications creating the worst conditions possible for delivering HDD based storage IOPS. Random IOPS per 15K RPM 600GB small form factor SAS Enterprise HDD will range from a couple hundred to rare cases of low thousands depending on IO size, mix of IO sizes from concurrent multiple applications, mix of reads and writes, as well as percentage of random IOs. The traditional workarounds of DRAM caching and HDD short stroking have proven to be inadequate and unsustainable for today s workloads. To provide enough memory caching to make a significant difference in HDD based storage systems for these IOPS demanding applications is cost prohibitive. The general caching rule of thumb is to provide at least 10% of the total capacity in cache. Doing this increases the probabilities of read cache hits to increase IOPS enough to make a difference. However, that means a storage system as small as 10TB would require at least 1TB of DRAM. DRAM is the most expensive form of storage and it is volatile (data is lost when power is turned off.) Non-volatile versions of DRAM (battery or supercapacitor based) are significantly more expensive. The other HDD workaround of short stroking (only placing data in the outer sectors of the HDD platters) has the appearance of low cost but with high complexity. Short stroking can increase IOPS by 4
placing data only in the outer sectors of the HDD. Short stroking can increase random IOPS by as much as a third, but it reduces HDD usable capacity by 50% and as much as 90%. That makes short stroking a lot more expensive than it appears by driving up the cost per GB and since most storage systems are drive limited, it also limits the total maximum addressable capacity. One more thing; HDD short stroking requires skill and experience. It is very easy to make errors that wipe out any performance gains while still reducing usable addressable capacity. HP 3PAR StoreServ Flash Storage satisfies the IOPS hungry mission-critical appetites. It does so by optimizing flash SSDs in a unique multi-controller mesh architecture that increases random IOPS by 20 to 40x over those provided by the fastest 15K RPM HDDs. Those gains expand even more when compared to 10K RPM HDDs. This enables HP 3PAR StoreServ to not just meet but also typically exceed IOPS hungry mission-critical application requirements with fewer drives, racks, and infrastructure. Yeah but Flash SSD Storage Mythology Inhibiting Adoption Storage administrators have always been somewhat risk averse. There is consistent resistance to new technologies and new ways of solving storage problems. The three most common objections to obtaining flash SSD storage are: 1. Reliability and Endurance Mostly this objection is the result of how data is stored and changed on Flash SSD stores is completely different from how data is stored and changed on HDDs. HDD technology originated more than 60 years ago. Ones and zeros are stored as magnetic + or polarity. That technology makes it quite simple to change or alter data in place simply by overwriting it. There is no wear on the HDD platters for writes or reads. Flash SSD technology is not at all similar in how data is stored and changed. Data is stored via electron capture in a NAND flash cell. The state of a cell is determined by its charge. The number of bits allowed per cell determines the number of states a cell can exhibit. For example: single level cells (SLC) allow one bit per cell enabling two states (0, 1); whereas enterprise or consumer grade multi-level cells (emlc or cmlc) allow two bits per cell enabling four states (00, 01, 10, 11). More bits per cell enable higher density but also require more power to store and read the data. More power causes more errors. These errors can be corrected in the flash SSD controller and to a greater extent in the Storage System controller. That s where HP 3PAR Flash Storage delivers industry-leading error correcting code (ECC). Data written to flash is stored in blocks ranging from 4K to 128K in size. Once any part of a block has captured data, no other data can be written to that block until the entire block is completely 5
erased. In other words data cannot change in place. It doesn t matter if only a small percentage of a block s capacity actually has captured data, no other data can be written to that block regardless of the unused capacity. Once erased it becomes available again to write data. Erasing a flash block destroys a layer of material. This is known as a program-erase cycle (P/E) colloquially known as wear. As data is aged out because it has changed (new versions of the data have been written on other blocks) the obsolete blocks are earmarked for erasure. This is known as garbage collection. Then the obsolete data blocks are erased and put back in the available pool for writes. Most Flash SSDs have a 10 to 25% additional capacity typically not noted on the drive to ensure a sufficient pool of available blocks for writes. To prevent any given block from being targeted too often for writes and wearing out prematurely, the flash SSD controllers provide wear-leveling over all of the blocks. Flash SSDs have a fixed wear-life measured as complete drive writes per day (DWPD). MLC flash SSDs DWPD typically range from.5 to 45 for up to 5 years. Small block sizes have less waste than larger ones but require the controller to keep tabs on a lot more blocks for garbage collection and erasure, slowing the SSD. Large blocks are much less efficient but there are fewer of them to track making the SSD faster. Flash SSDs or parts of flash SSDs can and do wear out. That is the concern with reliability and endurance How HP 3PAR StoreServ Flash Storage Makes Reliability & Endurance an Advantage HP 3PAR StoreServ Flash Storage has been architected to cleverly get the most possible reliability and durability out of its flash SSDs. It does this in several ways. Minimized wasted block space and P/E wear is delivered by the HP 3PAR StoreServ Flash Storage ability to coalesce writes in memory. This converts random IOs into sequential. Sequential IO eliminates or at least curtails wasted block space. Wear leveling is taken to unprecedented levels by providing it not just within a SSD but also across all of the SSDs in the system. Data writes are reduced in three ways. The first is HP s unique zero detect that does not write zeros to flash SSDs. Quite useful for VMware zero volumes. The second is high speed Inline deduplication that eliminates writes of duplicate data before it consumes flash SSD storage. And finally, the HP 3PAR Flash Storage eliminates extra writes also by detecting cache write hits. Flash SSD endurance is significantly enhanced as a result of HP 3PAR StoreServ Flash Storage distinctive Adaptive Sparing that unlocks the hidden capacity present in each flash SSD. Unlocking that hidden capacity permits the HP 3PAR StoreServ Flash Storage to utilize that spare capacity as additional over-provisioned capacity. This is what enables HP to convert a 6
SAS SanDisk 3.84TB read optimized flash (SSD) to being both write and read-optimized and guarantee a 5 year warranty if it ever wears out. HP 3PAR StoreServ Flash Storage series are the first all-flash storage systems providing write and read-optimized 3.84TB SSDs. T10 DIF protection To ensure data integrity, the 3PAR architecture computes DIF for every incoming IO and verifies the integrity of data once it gets completed by media including SSDs. This ensures that all transmission and data storage errors are caught and corrected end-to-end. 2. Density Density is both a matter of raw/usable capacity per drive, the number of drives that can be packed into a unit of rack space, and the total pragmatically addressable capacity (within the performance requirements) by the storage system. For several years now, the Fat HDD running at 7200 RPM has been the drive density leader, currently at 10TB max raw capacity per HDD in the general market. Fat HDDs utilize the large form factor of 3.5 that is approximately more than 50 times larger by volume and more than 14 times heavier in weight than the size of the small form factor (SFF) 2.5 SSD drive. To pack in as many as possible in the fewest number of rack units has inspired some clever enclosures allowing as many as 60 to 98 large form factor HDDs in as little as 4U (120TB to 167TB per RU). That s impressive density. Regrettably these fat drives are impractical and inappropriate for primary application workloads. Their performance per GB is much too slow (less than a third of that of SFF SAS HDDs per GB). And even serviceability of the HDDs is far too labor intensive (most of the high density rack units require server lifts and ladders to replace a single failed HDD) since the drive enclosures are often top-loading instead of frontloading. That means the 2.5 SFF high performance (15K and 10K) HDD density is a more appropriate comparison to SSDs. Both utilize the same form factor and enclosures so the comparison comes down to the raw and usable capacities of both. The flash NAND chip density gains from smaller die sizes and increased performance and durability of MLC from more sophisticated software have changed the game. How HP 3PAR StoreServ Flash Storage Density (Capacity) Blows Away HDD Storage Density Flash SSDs are based on microchips. Microchips are ties to Moore s law (number of transistors per chip doubles every couple of years.) Rapid die size shrinkage has accelerated SSD capacities and density. As previously discussed, the HP 3PAR StoreServ Flash Storage systems utilize SanDisk Optimus Max 3.84TB SSD that HP enhances from read optimized to both read and write optimized. The most cost effective high performance 7
SFF HDD is the 600GB raw capacity drive. That means there is about 640% more raw capacity per SSD and 1280% more usable capacity based on HP s conservative average data reduction/compaction techniques. When applying these two media types to an average configuration (based on HP market data) of approximately 115.2TB usable capacity, the numbers are even more lopsided towards the SSDs. The traditional high performance SFF HDDs requires 128 of them. Those HDDs require six 24- drive enclosures of 2U each for a total of 12 rack units. Those 128 HDDs equates into 76.8TB of raw capacity and 115.2 TB of usable capacity (using an average formatting and RAID overhead of approximately 25% and 2:1 data compaction from thin provisioning and zero block deduplication.) The comparable 3.84TB SSD configuration requires just 10 drives to get to 115.2TB usable capacity. The usable capacity obtained could well be higher depending on customer workloads but HP only uses a conservative 4:1 average data compaction rate on these SSDs. The resulting configuration consumes a single 2U 24-drive enclosure (< 17% of the rack space), and delivers 8 to 10 times the IOPS. These figures are even more lopsided in the HP 3PAR StoreServ Flash Storage 20800/20850. It is indisputable that HP 3PAR StoreServ Flash Storage SSD density significantly exceeds high performance 15K and 10K RPM HDDs by a considerable margin. What is somewhat surprising is that the HP 3PAR StoreServ Flash Storage density is starting to approach that of storage systems utilizing fat 7200-RPM nearline drive storage., This is a bit startling especially since there is no comparison on performance. However, the $/GB are not there yet even though it s moving in the right direction. Each HP 3PAR StoreServ Flash Storage system scales to significant new flash storage heights. A single 2U rack of the 3.84TB SSDs provides about 280TB of usable storage and the HP 3PAR StoreServ Flash Storage systems are capable of using multiples of them. That enables a HP 3PAR StoreServ Flash Storage 7450 to scale to approximately 5.5PB of usable storage; a HP 3PAR StoreServ Flash Storage 20800 or 20850 to scale to nearly 11.8PB of usable flash SSD storage. (Note: the 20800 can add roughly another 3.4PB of SFF HDD usable storage for a total of around 15PB of usable storage.) Neither flash SSD scalability nor flash density is an issue with HP 3PAR StoreServ Flash Storage systems. But the number one objection is cost. 3. Total Cost of Acquisition (TCA) and Total Cost of Ownership (TCO) The flash SSD market perception is that TCA is high and that TCO is even higher. This has led to the opinion of many IT pros that flash SSDs can only be justified for the most performance/iops demanding applications and nothing else, thus relegating flash SSDs to a niche. The metric utilized to describe both flash SSD and HDD TCA is generally $/GB. 8
There is an old saying that states: There is only one chance to make a first impression. The first flash SSDs were quite expensive. They were based on more costly SLC, but solved the performance problem for latency and IOPS hungry applications. It is interesting to note, that whereas TCA has been the consistent measure of cost whereas TCO has rarely been utilized. As noted throughout this document, flash SSDs have changed. Reliability, durability, and density have all increased by orders of magnitude and show no signs of slowing down. All of this has had enormous impact both TCA and TCO. How HP 3PAR StoreServ Flash Storage Now Costs Conspicuously Less Than Equivalent HDD Storage Using the previous discussed average configuration at typical street prices (including 3 years of Proactive Care 24 Support ) produces some astonishing differences in both TCA and TCO. The 15K SFF HDD configurations are coming in at roughly $2.12/Usable GB TCA whereas the 3.84TB SSD configurations are coming in at approximately 22% less at $1.65/Usable GB TCA. What s even more remarkable is that the 3.84TB SSD configurations are coming in approximately 10% less than the performance 10K SFF HDD configurations at roughly $1.84/Usable GB TCA. In maximum 3.84TB HP 3PAR Flash Storage configurations, that cost per usable/gb TCA declines to less than $1.50, which is approximately 20% less than 10K HDD configurations. This impressive cost gap grows significantly when power and cooling are added into the picture. The HP 3PAR StoreServ Flash Storage SSD annual power and cooling consumption averages generally 10% per usable GB of equivalent SFF HDD storage. Since power costs vary widely by region, for the purposes of this document, the rule of thumb utilized for calculating HDD power costs will be one times purchase cost per year. This is an imperfect comparison and typically utilizes a lower power and cooling cost than actual; however, for comparative purposes it will demonstrate how HP 3PAR StoreServ Flash Storage is in fact much lower in cost than equivalent HDDs. Taking the TCA plus power and cooling costs produces some mind changing number. The 3.84 SSD configurations comes in at a whopping 73% less than equivalent 15K HDDs and 69% less that the equivalent 10K HDDs. Note this is not a total TCO and those gaps will grow still larger when other cost considerations such as rack unit space are taken into consideration. What this means is that HP 3PAR StoreServ Flash Storage usable capacity costs considerably less than equivalent HDD capacity. What All This Means In The Real World All data center workloads can now run faster, with higher reliability, and lower costs. Batch workloads previously relegated to HDDs that took hours to complete can now be completed in minutes. Time windows that were being compressed by the needs of the business are no longer a problem. Maintenance tasks that get continually put off can be completed. Application 9
development is accelerated. Database replications no longer take too long. Everything happens much faster with higher reliability and lower cost. This is all made possible by the HP 3PAR StoreServ Flash Storage systems utilizing the SanDisk 3.84TB drives. This combination delivers significant advantages over 10K or 15K RPM HDDs with: Up to 40x better performance; Greater reliability; More than 3x greater density/capacity; And best of all 10 to 20% lower cost per usable GB. The one non-renewable resource in the life of every IT administrator is time. The HP 3PAR StoreServ Flash Storage systems give IT admins more of that priceless precious time at a lower cost. For More Information Go to: www.hp.com/3par 4AA5-9954ENW 10