Guest Post

Storage Processors to Unleash the Full Potential of Flash Storage

It is time to simplify infrastructure, stop the over-investment in storage and scale more efficiently. Here are insights to help you unleash the full potential of your flash storage investment.

Download the presentation: Unleash the full potential of flash storage

00:04 Steve Fingerhut: Thank you for joining us today. I'm Steve Fingerhut, president and chief business officer for Pliops. I'm excited to help you unleash your flash storage for the data-centric era. To introduce Pliops, please meet Uri Beitler, CEO.

00:21 Uri Beitler: Thank you, Steve. When we founded Pliops, we carefully analyzed the key data center trends emerging in the triangle of compute, network and storage, and realized that they are on a collision course. Those trends are accelerating today, limiting the scalability of data centers and resulting in excessive cost of billions of dollars in inefficient infrastructure. Pliops has developed a revolutionary storage processor that bridges this gap for today's and next-generation data center enabling the . . . evolution.

The first key trend is flash adoption, which is taking off with six times more flash used in data centers today than in 2015, but this is just the tip of the iceberg as the exponential growth continues and expected to grow over 40 times until 2030, accounting for up to 50% of the bit stored, but such capacity is just part of the equation as the performance levels provided by NVMe SSDs are over three orders of magnitude faster than HDDs.

01:26 UB: While we're experiencing the huge growth in data capacity and performance, the applications are becoming data driven as well and require more and more data. However, general purpose CPUs are stagnating with the decline in Moore's Law and failing to provide the needed performance. Basic accelerators already exist for analytics and I/O flash arrays, and in Pliops we're taking it to the next level and invented ultrafast key-value acceleration engine as the foundation to storage processing, that eliminates the redundancies in the storage and database stack. Live storage processor is a key component assisting both existing and emerging storage system architecture that attempt to solve the major data center challenges. We have prepared a short video to explain the source of these challenges.

02:16 Speaker 3 (video): Applications such as SQL and NoSQL databases, use key-value storage engines that enable variable sized objects. While not computationally challenging in a hard drive, an SSD capable of 500K IOPS is a different story. With the expense of flash and its endurance write limitations, there is strong motivation to compress flash resident data, but both the compression and managing that data is computationally expensive. By offloading key-value storage to an optimal architecture on a dedicated hardware device, Pliops Storage Processor reduces inefficiencies and improves performance by cutting CPU utilization to a small percentage of what typical software solutions demand. Pliops accelerates compute-intensive functions and eliminates bottlenecks to put an end to wasted storage and allow you to take full advantage of your flash storage investments.

03:14 UB: These challenges are not going away, and we need to solve them now because they cause major inefficiencies across the stack. We see this today with storage sitting idle while possessors are maxed out, and in order to keep up with performance, software takes shortcuts that amplifies data read, written and stored up to two orders of magnitude. And many avoid usage of advanced data protection and reduction schemes due to the performance penalty.

Another example that we often hear from our customers is that storage is over-provisioned extensively in order to overcome performance issues, not to mention the excessive cost and maintenance for servers, which we offload as well to our processor.

03:53 UB: Let's work through how this multiplies. Start with your data, then add the fragmentation due to sub-optimal packing and merging. Another layer of redundancy comes due to storage policies that require room for data growth at the volume level, then duplicate that data for A10 and you easily get up to six times more physical storage than data, even before taking into account mirroring and duplications between nodes. Naturally, storage architects and their CFOs are demanding a better way. It is time to unleash the full potential of your SSD storage and Pliops is here to make it happen.

The Pliops Storage Processor optimizes the way data is being stored, managed, accessed, analyzed and protected, resulting in significant value of performance, cost, endurance and power consumption. Pliops Storage Processor provide generic value across the broad range of applications with one-cloud-fits-all acceleration. This makes it valuable and easy to deploy across your data center. Steve will explain in more details the storage processor. Take it away, Steve.

05:02 SF: Thanks Uri. This is a mirror of what Uri presented earlier, we start with the inefficient, bloated SSD configuration and Pliops data protection eliminates the waste from mirroring and brings advanced in provisioning to reduce the layers of safety margin and over-allocation, especially in multi-tenant environments. And finally, Pliops inline transparent compression reduces the data to a fraction on disc.

Sound like magic? It's close, except it's real. Let's look under the hood. The core innovation is a hardware-based key-value store. Key-value stores are used in nearly all database, analytics and software-defined storage applications, it is complex software to manage what we saw in the earlier animation: compress, pack and merge, garbage collect. This is very complex and inefficient.

05:54 SF: Pliops has developed breakthrough data structures and algorithms that deliver the equivalent of hundreds of cores of host software. Accessible via block interface, this literally appears as a block device in the system, meaning this will accelerate any application. To cut through decades of legacy software, Pliops offers a direct KV interface that uses the same APIs as your software. It's an easy change to supercharge your software. We've worked with partners to do this for several applications and it is easy, the KV interface uses the same APIs as software. Let's see how this looks in action.

06:31 S3: Key-value stores are used in nearly all database, analytics and software-defined storage applications. CPU and system resources are over-utilized when running this complex software. This results in applications not taking full advantage of SSD storage capability. The picture dramatically improves when offloading to Pliops Storage Processor with an ultra-efficient KV engine. This will enable system resources to scale with more capacity available from your SSDs, while the data is protected and encrypted. Pliops Storage Processor delivers a significantly higher performance, allowing more applications to run on the host.

07:15 SF: Pliops Storage Processor is also easy to deploy in your preferred system. Our low-profile card fits in any standard server. And it works with any SSD, direct attached or disaggregated. There are no special hardware requirements like peer-to-peer support or direct cabling. Easy. And the performance is off the charts. On core block I/Os, Pliops Storage Processor delivers massive performance and handles random writes with ease. With databases like MySQL, MongoDB and Cassandra, Pliops delivers major increases in instance density while reducing latency, so you can push the system to the max and still deliver a great user experience.

We have a special guest, Peter Zaitsev, CEO of Percona. Peter, what are your thoughts on Pliops Storage Processor?

08:07 Peter Zaitsev: For, now, more than a decade, we have been helping many of the leading enterprises and internet big names to be successful with their open source databases. One thing which has been happening in the industry, and continues to happen, is absolute explosion of the data. And that means there is a lot of concerns about being able to handle this data with a good performance and a great efficiency. And that is, I think, where Pliops innovations are absolutely fantastic.

Now, over the last decade or so, we have seen the transition from their spinning hard drives to their flash and other types of solid-state storage, right? And that was a transition which provided huge performance gains, right? But a lot of the that has since leveled off, because we still have been treating those storage devices as dumb block devices, which is not really very helpful for database workloads. And I think this is a great innovation, which Pliops brings there, instead of flash storage being dumb, it is actually smart and it's able to take some of the database processing needs on its own with providing their key-value interface, which database can use instead of a dumb block storage interface.

10:11 PZ: If you think what concerns one may have with the database, specifically on a flash storage, I see that goes into their three different buckets. One, is of course performance, response time, that is always what database users are concerned about. The second, is space utilization, because flash tends to be expensive. And the third, which is specific to flash, is the write endurance or write amplification, because the more writes flash is able to handle, the more expensive it's going to be.

And what is unique for Pliops Storage Processor is that it is able to help with all of those three components. It's, according to our benchmarks, it provides substantially better performance, with increased throughputs and lower latency, it provides a better compression, the native MySQL solutions reducing the storage requirements, as well as its reduced write amplification, that means the same storage will last longer, or you can put heavy write workloads on the same storage without reducing its lifetime.

11:44 SF: Thanks, Peter. Redis is an extremely popular in-memory data service. The challenge is, it requires DRAM for all data, which is quite expensive. We compared performance with DRAM, and per system, Redis delivers 900,000 IOPS and one millisecond four-nines latency. With flash versions, the performance drops 90% and latency grows nine times.

Pliops, using low-cost data center SSDs, matches the performance and latency of DRAM at one-seventh the cost. Pliops enables a 10-1 server consolidation by taking advantage of Pliops' high performance and the high capacities of flash. And this is not the result of caching. In the tests used, 90% of the I/Os were served from flash. Customers are interested to use Pliops to expand, from accessing a few terabytes of data to hundreds of terabytes at in-memory speeds. But don't just take our word for it, please welcome to Takeshi Hasegawa, deputy manager of product development at Sakura, the largest hosting service in Japan.

12:51 Takeshi Hasegawa: Hi, I am Takeshi Hasegawa at Sakura Internet. I had been a sales engineer at Fusion-io before as my background. Today, I would like to share what I have found through the evaluation of Pliops Storage Controller. As a cloud computing provider, it's no doubt, flash cost will take major portion of the system. Pliops would allow more utilization and performance for the investment, supporting more concurrent users with predictable response from identical drives.

13:22 TH: The data is transparently compressed, so we will have additional usable capacity. It also reduces write amplification, the back-end flash lasts longer. Some of our customers have better servers with NVMe flash from us, but they still want to push more database performance. What could be upgraded for them?

Pliops Device Processor is, according to my benchmark results, thanks to the superior architecture even bests RAID 5 perform better than typical RAID 0 configuration. I have never seen such results for the rest of 10 years ago outside flash. Sakura has an affiliated software-defined storage company, and we are in the discussion of reverting the Pliops. We see the benefits mentioned before and we are also examining the key-value store API interface from Pliops. Once we move our primitives bypassing the six file system and broke interface, be able to get a huge performance improvement on metadata processing. I'm also personally enjoying the time to work with the Pliops team and looking forward to the deployment. Thank you.

14:37 SF: Thanks Hasegawa san. We're just getting started. A key benefit of Pliops Storage Processor is it reduces writes to the SSD by up to 90%. For QLC, the key challenge is their limited random write performance and write endurance. With a 70/30 workload, Pliops delivers 18 times the performance of software only, whether it's TLC or QLC will perform the same with Pliops and increases endurance up to 10 times. Pliops can also take advantage of the lowest cost Zoned Namespaces in QLC SSDs with no software changes required. Optane has amazing performance and latency, but Pliops not only improves performance and reduces latency, it greatly increases the performance per core, enabling more scaling for your investment.

15:34 SF: All of these results are with high performance inline compression, so you can store up to four times the data on Optane reducing the dollar per GB by 75% or more. Pliops' innovative architecture enables drive failure protection that is breakthrough in performance, eliminating the trade-offs that exist today. No hot spares required, so you get the full performance and endurance of your investment.

Another barrier to RAID 5 with SSDs is the huge hit in performance when a drive fails in degraded mode and during rebuild. Pliops resolves that, maintaining very high performance during drive failures and rebuilds.

16:13 SF: This makes drive failure protection relevant if you were using RAID 10 or even if you've given up and use RAID 0. Distributed storage systems also benefit by improving reliability and eliminating network rebuild traffic from failed drives.

The worst part of RAID 5 is performance during degraded mode, and while rebuilding a failed drive. Not only is storage performance much lower, the CPU is busy, which takes away from host applications. Pliops virtually eliminates these trade-offs, so storage performance remains high, and you have the full performance of the CPU.

Now, I am pleased to welcome our chairman, Aryeh Mergi. He was the co-founder of M-Systems, which is the inventor of SSDs as well as XtremIO, a pioneer in all-flash arrays. Two years ago, Aryeh received the Flash Memory Summit Lifetime Achievement Award for his work in breakthrough innovations that helped to bring the flash storage industry to where it is today.

17:15 Aryeh Mergi: I hope that you are as excited as I am about the amazing technology and value the Pliops team showed you today. At M-Systems and XtremIO, we invented key enabling SSD technologies and were the first to bring all-flash arrays to data centers. In their own way, each brought the flash industry to new levels many in the industry didn't think possible.

In the first years of flash, the industry was busy solving the challenges of making flash a great and faster replacement of hard drives in existing and new applications. With the advance of flash and the exploding needs for speed and densities, flash is no longer the bottleneck. CPUs and hard drives derive block API are now stopping flash from delivering to their capabilities.

18:12 AM: Today with Pliops, we are breaking through those bottlenecks and have developed new storage compute acceleration technologies, which are enabling new levels of performance, and which we believe will be a key component of all next generation storage systems. Pliops Storage Processor shares many of the same values as the SSD and all-flash array at its beginning. A focus on ease of use and delivering breakthrough levels of customer benefits. I invite you to join us on the journey to unleash the power of flash storage for the data-centric era.

18:52 SF: There's more to learn. Attend our sessions. Visit us at the virtual booth and go to pliops.com for more information. Thank you.

Dig Deeper on Flash memory

SearchDisasterRecovery
SearchDataBackup
SearchConvergedInfrastructure
Close