LeMagIT: Graid Boosts Performance of NVMe RAID with a…GPU

(LeMagIT.fr, Yann Serra, October 14, 2022)

“We rebuild in two hours what an ordinary RAID system takes three weeks to rebuild.”

Tom Paquette // Managing Director, GRAID

An Nvidia graphics card as a RAID controller! This surprising idea is the basis of the “SupremeRAID™️” solution from the American startup GRAID. For less than $4,000, it can drive 32 NVMe SSDs simultaneously in a server to achieve a total read throughput of 110 GB/s and an access speed of 19 million inputs-outputs per second (millions of data). ‘IOPS).

By comparison, Broadcom’s LSI MegaRAID card, its closest competitor, directly supports only four NVMe SSDs, reads their data at a maximum throughput of 13.5 GB/s, and delivers less than 200,000 read IOPS. The difference between the two cards? Broadcom uses an ASIC, a custom chip that the manufacturer has developed itself.

“Obviously, we started by trying to create a RAID card from a specialized chip, in this case, an FPGA. But we quickly became disillusioned: to reach the 19 million IOPS we were aiming for, our FPGA had to cost $30,000 each! It made no sense,” says Leander Yu, CEO of startup GRAID.

“In contrast. Knowing that RAID is just to distribute the accesses in parallel streams and that a GPU precisely has the function of parallelizing the streams of pixels, we had the idea of ​​taking a simple graphics card available everywhere on the market for a few hundred dollars,” he adds.

Thus, the SupremeRAID  SR-1010 controller card is really just an NVidia RTX A2000, an entry-level graphics card for workstations , which can be found nowadays for less than 600 €. “Honestly, we don’t even use 50% of its power. On the other hand, we had to choose a graphics card of professional class. Consumer models do not have error-correcting memory. They are less reliable. We cannot tolerate the loss of a single byte in storage,” says Leander Yu.

The SupremeRAID SR-1010 is the new model on PCIe 4.0 bus.  Graid already marketed last year an  SR-1000 model with PCIe 3.0 bus and which reached 16 million read IOPS. The SR-1010 is priced at $3,995. The older SR-1000 costs $2,500.

Leander Yu adds that, according to him, the solution of its competitor Broadcom was certainly developed at a time when hard disks had to be RAIDed. It would be obsolete with NVMe SSDs, which no longer communicate via SATA or SAS ports , but now use PCIe channels. He believes that Broadcom probably kept the same ASIC to make profitable its very expensive development.

12 times faster on writes

The technical know-how of Graid is really only in the software implementation on GPU of RAID 0, 1, 5, 6, 10 and EC modes. “We sell the complete solution, with the card. But we are also in discussions with server manufacturers and cloud hosts who are buying very large quantities of graphics cards. We are open to selling them only our software,” says Tom Paquette, GRAID’s general manager.

In practice, the card only serves as a gateway to the SSDs during writes. Because that’s when it has to take the streams and segment them into fragments so that they can then be played back in parallel with maximum speed. In this configuration, the write throughput from the server’s point of view is 12 GB/s, which is the maximum speed of the PCIe 4.0 x 16 channel bus on which the SupremeRAID card is connected.

On output, the card communicates writes to each SSD at the individual speed of 7 GB/s, their maximum speed on a PCIe 4.0 bus when they use a maximum of 4 channels. But if we count the overall throughput to all SSDs, it comes to 22 GB/s, due to the duplicates that are created by the RAID device.

Graid compares its solution to “a competing hardware RAID controller” – without specifically quoting Broadcom this time – which can only write a total of 4 GB/s to SSDs. Obviously, because its chip had been designed for hard drives, therefore, and not for NVMe SSDs. A software RAID is also compared, in which it is the server’s processor – which is also busy running the applications – which is responsible for writing to the SSDs. This time, the speed would drop to 2 GB/s in writing.

In terms of access speed, the SupremeRAID would reach 1.5 million write IOPS, against 180,000 IOPS for the “hardware controller” and 200,000 IOPS for the processor alone.

No read bottleneck

It is on the other hand on the readings that the solution is particularly effective. “The essential point of our solution is that it does not communicate to the server the data that it must load. It tells it which SSDs to load them on and the server reads them directly from several SSDs in parallel,” explains Tom Paquette. This is how the record speed of 110 GB/s would be reached on a server equipped with two Xeon 6338, 32 cores each, at 2 GHz.

Opposite, Broadcom’s solution continues to act as a gateway when data is read from SSDs. That is, even giving it a PCIe switch that would allow it to be intermittently connected to 32 SSDs, data would still only have its 16 PCIe channels to flow back up to the server’s RAM.

And Tom Paquette adds: “we are not limited to internal SSDs. Our controller can communicate very well with an NVMe-over-Fabric controller installed in another PCIe slot and driving SSDs in an external drawer. That is, in an extreme use case, we can do RAID that writes and mirrors all data to both internal SSDs and external SSDs. We are in fact the only ones to offer NVMe-over-Fabrics compatible RAID,” says the CEO of GRAID.

He specifies that the limit of 32 SSDs is not due to the capacity of the Nvidia card, but to the software of GRAID. “These amounts will improve as we update and machines are able to support so many SSDs. Nevertheless, our software is already planned to be able to be used with the future standard CXL of composable architecture”.

The CXL bus architecture is the one that will eventually extend the next PCIe 5.0 buses, so that it will become possible to install processors, memory modules and PCIe extensions in physically separate nodes. This type of architecture is supposed to offer more possibilities for assembling clusters of machines.

GPU Speed ​​for RAID Rebuilds

Another very important point in RAID: reconstruction times, which are usually prohibitive. They arise when a disk fails and all spare duplicates must be referenced to make them usable blocks of data. It is in this area, which usually requires a lot of computing power, that the GPU of the graphics card shows its full potential.

“It’s very simple, we rebuild in two hours what an ordinary RAID system takes three weeks to rebuild,” boasts Tom Paquette, without however associating these delays with a capacity.

According to more precise measurements, the SupremeRAID card of Graid would allow to have an access speed of 5.5 million IOPS in reading and 1.1 million IOPS in writing during the reconstruction phase. While Broadcom’s MegaRAID would drop to 36,000 and 18,000 IOPS respectively.

“In this regard, we also evaluated Nvidia’s Bluefield DPU. But these chips are more designed to encode and decode on-the-fly data circulating in the network. If we didn’t choose them, it’s precise because they don’t have enough power for the functions that interest us, typically RAID reconstruction”, indicates Leander Yu. LeMagIT asked him about the relevance of using a graphics card, while NVIDIA has invested heavily in the development of accelerator chips to communicate with external storage .

For machine learning and continuous video recording

Among the use cases of the SupremeRAID card, Graid first cites machine learning engines that must constantly load very large volumes of data into memory. According to the startup, an algorithm that usually takes 12 hours to run with data stored on a NAS in NFS, would only take 2 hours by communicating in NVMe/RoCe with a storage server equipped with the startup’s controller. On average, a storage array offers an access speed of 1.5 million IPOS, which is 12 times less than a server equipped with SupremeRAID.

Another example, the video recording of a car race which requires a bandwidth of 10 GB/s between the cameras and the storage, could finally benefit from RAID 5 or RAID 6 data protection (calculated distribution of duplicates). Another configuration might only offer RAID 1, which doesn’t require as much computing, but would consume 80% more storage space.

“Enterprises are increasingly buying NVMe SSDs for performance, but they find themselves in the paradoxical situation where this performance is impacting application performance. Because the processors have fewer cycles to grant them that they have to redouble their calculations to manage SSDs in RAID. We deal with this paradox”, argues Tom Plaquette.

Soon PCIe 5.0 and Erasure coding

These days, the software of the SupremeRAID is evolving into a 1.3 version which supports more Linux distributions. It also becomes possible to install two controller cards in a server, for redundancy (and not to double performance or the number of disks).

In 2023, the software will be equipped with a graphical interface to administer the card under Windows, Erasure coding which will allow RAID to be made between several cards, as well as support for the next PCIe 5.0 buses. The roadmap cites compatibility with Kubernetes and VMware in the more or less short term, but also compression, encryption, and on-the-fly deduplication.

“This is another advantage we have over other RAID controllers: we can upgrade our solution every year, whereas our competitors wait at least three years before being able to update their card”, concludes Tom Paquette.

---

About Graid Technology

Chosen by CRN as one of the Ten Hottest Data Storage Startups of 2021 and a 2022 Emerging Vendor in the Storage & Disaster Recovery category, Graid Technology Inc. has developed the world’s first NVMe and NVMeoF RAID card to unlock the full potential of enterprise SSD performance. We’re headquartered in Silicon Valley, with an R&D center in Taiwan, and are led by a dedicated team of experts with decades of experience in the SDS, ASIC and storage industries. Graid Technology Inc. is redefining performance standards for enterprise data protection: a single SupremeRAID™ card delivers 19 million IOPS and 110GB/s of throughput. For more information on Graid Technology Inc., connect with us on Twitter or LinkedIn.

Additional Resources

• Blocks & Files Compares Graid Technology’s Killer RAID Against The Competition
• Graid Technology: The Future of Data Performance – Silicon.de
• GPUs to Accelerate NVMe Storage in RAID Mode – Lemonde Informatique
• Information Age Discusses The Technical Success of SupremeRAID™️