Intel SSD DC P4510 drives promise dense, efficient bulk storage

Intel expects that change is coming in the data center. The way the company sees it today, the various parts within today’s typical server—CPUs, SSDs or hard drives, memory, and expansion cards—become functionally obsolete at different points within the server’s lifespan, ultimately leading to a large cash outlay to replace the entire box at its end of life. Contrast that with the smaller bills involved with replacing the parts that are no longer pulling their weight while keeping what works while the server is still in use.

Racks filled with “converged” servers—those with memory, storage, and computing resources inside—also might employ their storage and computing resources unevenly. The storage capacity within a server or servers might be maxed out, for example, while portions of the compute resources in that same server might sit idle. The bandwidth and latency characteristics of NAND flash could also prove the ultimate bottleneck to getting the most performance out of each server in a rack, as well.

At least for buyers of its storage products, Intel thinks it sees a better way of building up future servers and racks. Instead of continuing to build up racks with today’s idea of a converged server, the company believes it can help businesses get more out of each server’s compute resources by replacing NAND SSDs with pools of Optane DIMMs and Optane SSDs for “hot” data. The impressive low-latency characteristics and high capacity of Optane devices could lead to as much as 18 TB of virtual memory per server for applications that need both speed and capacity. Meanwhile, NAND flash would be relegated to its own rack units for bulk data storage instead of taking up valuable space within those servers.

The SSD DC P4510 that Intel is unveiling today is meant to be one of the building blocks for those bulk-storage tasks. These drives use Intel’s latest 64-layer NAND flash chips to boost both density and performance, and they also benefit from a focus on quality-of-service improvements. The company says it’s been shipping tens of thousands of these drives to cloud service providers since the third quarter of last year, and it’s now ready to begin shipping them in 4-TB and 8-TB capacities on top of 1-TB and 2-TB models.

Although the DC P4510 offers a claimed 90% increase in raw write bandwidth over its DC P4500 predecessor (for 128K pages at QD128, anyway), that performance increase alone isn’t really the important part of this drive’s story for its intended audience of data centers and cloud service providers. The bigger improvements are for quality of service and power efficiency. Intel claims as much as a 2x performance-per-watt improvement from the DC P4500 for sequential-write workloads.

On the quality-of-service front, Intel improved the firmware in the P4510 to better prioritize reads over writes when needed and at a finer granularity than that of its past drives. The drive’s denser NAND doesn’t only allow for higher-capacity SSDs—it also offers lower program latency than the 32-layer stuff in the DC P4500.

Versus the older drives, those improvements cut the DC P4510’s read latency at QD1 for four nines (or for 99.99% of IOPS) from 2.24 ms to 0.17 ms and from 4.1 ms to 1.82 ms for a mixed workload of 70% random reads and 30% random writes. The company touted improvements out to six nines (or the 99.9999th percentile), and even at that demanding standard of QoS, the P4510 still delivers a huge improvement over the DC P4500.

Across a range of workloads, Intel believes the P4510 will offer a considerable speedup compared to the P4500. 

Intel’s results with the Aerospike certification tool for NoSQL workloads

In a test performed with the Aerospike Certification Tool, which measures a drive’s performance as a multiple of a baseline workload of 2000 reads per second and 1000 writes per second and only passes a drive if it services 95% of those IOPS within a millisecond, the DC P4500 offers a 4x result and the DC P4510 turns in a massive 48x score. For analytics workloads, Intel created an internal version of the STAC-M3 benchmark that shows the DC P4510 leading the P4500 in 15 out of 17 of the benchmark tests. In Intel’s favorite “write-pressure” torture test of SSD performance, the DC P4510 delivered 3.75x the bandwidth of the P4500.

The density and power-efficiency improvements in the DC P4510 could free up more space and power in a rack for other types of resources, too. Going by the Open Compute Project standard of 12 kilowatts of power per rack and giving over no more than 50% of rack space to storage devices, Intel offers an example of how one can get to 720 TB of storage within those constraints. WIth the company’s 2 TB P3520 drives, one would have to give over half the rack to storage units, and the total power consumption of those drives would add up to 4.3 kW—a little over a third of the rack’s power budget. Using 8 TB DC P4510s, on the other hand, one would only have to give six rack units over to storage, and those drives would only consume 1.4 kW.

The DC P4510 also improves a number of management and servicing scenarios versus older Intel drives. The P4510 works with the NVMe Management Interface spec, has online firmware update capabilities, and offers namespace management support for shared-storage environtments. On top of those features, Intel says it’s standardized the handling of surprise hot-plug-and-insertion events on its Xeon Scalable processor platforms with the DC P4510 in its U.2 form. The company says its older PCIe SSDs required vendor-specific implementations of hot-plug support that were apparently the cause of some headaches. This platform-standard approach comes thanks to a new on-chip unit called the Volume Management Device, or VMD. The VMD also offers a standard method of managing LED indicators to help data center administrators quickly locate and replace failing SSDs.

The VMD isn’t the only storage-specific feature baked into Xeon Scalable processors, of course. Virtual RAID on CPU, or VROC, arrived last year with Xeon Scalable processors as well, and Intel thinks it’s an ideal complement for DC P4510 SSDs. Intel notes that on its older platforms, customers had to turn to relatively expensive third-party RAID cards to gang together large numbers of SATA SSDs, making the RAID card the bottleneck when the CPU was hungry for data.

The costs of purchasing new platforms aside, Intel claims that using a compatible Xeon Scalable CPU, a $249 VROC key, and four DC P4510 4 TB drives can deliver as much as an 8X increase in IOPS using its test workloads compared to a $700+ discrete LSI 3108 RAID card and 12 Intel S35120 drives. If that’s true, VROC could offer a route to higher performance at similar or lower costs. Server OEM Huawei says four DC P4510 drives paired with VROC can deliver 500K IOPS on pure random write workloads in RAID 0, an apparently unprecedented degree of actual versus theoretical performance for such a configuration.

While the DC P4510 is launching in U.2 housings for the moment, Intel thinks the real future of its 3D NAND and the DC P4510 lies in the EDSFF, or Enterprise and Data Center SSD Form Factor. Born out of Intel’s Ruler SSD concept from last year, EDSFF lets Intel pack a claimed petabyte of SSD storage in just one rack unit, and the company says it will be able to deploy that density to customers this year. Density isn’t the only advantage of EDSFF, however, as Intel says its internal tests of cooling requirements for these drives let it cut airflow—and therefore cooling power usage—by as much as 55% compared to U.2.

All told, the DC P4510 seems poised to serve the needs of cloud service providers and businesses that need massive PCIe SSDs for massive data sets, with their attendant benefits for reliability, power usage, and heat generation. Data-center administrators who want to take advantage of the DC P4510’s virtues will likely need to get in touch with server OEMs or channel providers to get pricing and availability info, as Intel only claims that the drive is available now.

Comments closed
    • Bauxite
    • 2 years ago

    Shots fired at Broadcom (LSI, Avago), Adaptec, Marvell, Highpoint, etc.

    Although they had to see this coming, Intel these days only has pcie lanes for other vendors to use because it has to, not because it really wants to.

    • odizzido
    • 2 years ago

    until it bricks itself. You missed that in the title.

    • cmrcmk
    • 2 years ago

    So Intel’s concept is to run storage out of a separate shelf than compute and memory? Isn’t that how it’s been done in datacenters for decades? I run a Dell servers attached over Fibre Channel to a Nimble storage array already so I’m basically from the future, huh?

    (Hyper)converged has only been in vogue for a few years. Most companies haven’t yet had a chance to transition to it or have already turned it down for the exact reasons Intel has “discovered”: putting everything into one box makes it hard to maintain the right balance of resources for your workload.

    All that aside: the P4510 is sexy and I want it.

      • Chrispy_
      • 2 years ago

      I was just scratching my head thinking about this too. We moved to SANs as a company well over a decade ago.

      I run a mix of SANs (Nimbles, an EMC, a Nexsan, and a Supermicro homebrew NAS on steroids) It’s all running on a dedicated 10GBe iSCSI switched fabric and no servers use local storage other than for their hypervisor.

      Not only is that a common arrangement these days, it’s how Microsoft, VMware etc have been recommending you do things since about Server 2003.

        • willmore
        • 2 years ago

        I remember back in the late 90’s getting a grilling from the datacenter manager because we wanted local storage on our box rather than putting it on the SAN. We had spec’ed 8x 15K RPM drives in it which were to be an a RAID1–we really needed low read latency and we had done the testing to show that this config was what would work.

        So, yeah, I can’t see how they’re trying to pass this off as a new thing. Wasn’t EMC basically built on this architecture?

          • psuedonymous
          • 2 years ago

          I think Intel’s idea is that “we need large amounts of low latency local storage” is becoming a more common requirement for many workloads (especially as HPC encroaches further into the datacentre), with Optane allowing that to be clawed back out of the local boxes again.

        • Coldfirex
        • 2 years ago

        I see Nimble. You both get upvotes.

          • Chrispy_
          • 2 years ago

          Do you work for Nimble, or do you concur that they’re excellent flash-fronted bulk-storage?

            • cmrcmk
            • 2 years ago

            We’ve been running a Nimble array for a couple years now and I’m still giddy about it. We came from an IBM array that was a “great deal” but became useless once we started pushing virtualization. Now we’ve got 99% of our data on Nimble and it doesn’t break a sweat.

            5 out of 5. Would buy again.

    • willmore
    • 2 years ago

    You’re either unclear or misunderstanding ‘9’s notation. For example: 99.99% is four nines. 99.999% is five nines.

    For reference: [url<]https://en.wikipedia.org/wiki/High_availability#%22Nines%22[/url<]

      • Jeff Kampman
      • 2 years ago

      Whoops, corrected.

        • willmore
        • 2 years ago

        Thank you!

Pin It on Pinterest

Share This