Patrick Kennedy, Editor in Chief at ServeTheHome, has just published an in-depth, hands-on article and YouTube video about his recent Infrastructure Processing Unit (IPU) adventures at the Intel networking lab in Santa Clara, California. This was Patrick’s first hands-on experience with an FPGA-based IPU. The card that he worked with is a Silicom C5010X FPGA IPU, based on the Intel IPU C5000X-PL Platform. This platform incorporates an Intel® Xeon®-D 1612 CPU to manage the IPU’s control plane, and an Intel® Stratix® 10 FPGA to implement the IPU’s data plane. (Note: Intel sponsored Patrick’s trip to Santa Clara but had no editorial control over the subsequent article or YouTube video.)
In the lab, Patrick participated in running IPU experiments using a rack full of Supermicro Ultra servers serving as hosts and another Supermicro Ultra Server with eight 1.6-Tbyte Intel solid state drive (SSD) PC P4610 Flash-based storage drives serving as a storage target. It’s important to note that these SSDs are not configured as NVMe-over-Fabric (NVMeoF) drives with the accompanying NVMeoF software stack overhead imposed on the server. Because the IPU is plugged into the server’s PCIe bus directly, it’s able to make the remote SSDs look like local NVMe drives, even though they’re connected over 100 Gbps Ethernet links. In effect, the IPU offloads the NVME-over-Fabric stack and therefore lowers server CPU utilization and frees the server CPUs for more revenue-generating work. Patrick’s article and video about these experiments neatly encapsulate the essence of the arguments in favor of IPUs.
During the video, Patrick discovers how IPU’s user interface makes it simple to provision the SSDs and assign them to a server. The Intel Xeon-D processor on the IPU C5000X-PL Platform manages this control-plane interface. As Patrick observes, the IPU allows an ISP or cloud provider to quickly provision bare-metal servers, network bandwidth, storage, compression, and encryption on a customer-by-customer basis. The use of IPUs to manage servers in this manner allows data centers to pool storage so that any server in the data center can access this storage over Ethernet without performance bottlenecks. This implementation prevents storage drives from being stranded on a single server’s domain where it cannot be used by other servers, as is common in large data centers that lack IPUs.
Patrick also notes that the Intel Stratix 10 FPGA in the IPU C5000X-PL Platform can implement standard or custom encryption or compression algorithms (or nearly any other data-manipulation algorithm for that matter). The Intel FPGA’s capabilities deliver two benefits. First, the IPU’s onboard FPGA further lowers server CPU utilization by moving these tasks (encryption, compression, etc.) from the control plane to the data plane. Second, the data-plane FPGA accelerates these tasks because the hardware-programmable FPGAs perform these functions quickly, at wire speed. Because the FPGA is already in the data plane, data does not need to be transported from the data plane to the control plane for processing and then moved back to the data plane. This results in greatly improved efficiency, reduced server CPU usage, and reduced network overhead within the data center.
Further, Patrick says, it’s possible for an ISP or cloud service provider to develop custom compression or encryption algorithms and implement them in the IPU’s FPGA. There are at least two benefits to be gained from custom algorithms: security and differentiation. If unencrypted data never travels over an Ethernet connection, even within the data center, it’s more secure. If the Intel FPGA IPUs are used to implement proprietary algorithms that are even faster than industry-standard algorithms, there’s even more performance gains to be realized, further differentiating the data center’s offerings.
You will find some IPU performance results in Patrick’s video and article. We won’t spoil the ending but suffice it to say that Patrick was duly impressed with the results achieved by the IPU C5000X-PL Platform. As he writes in his summary of the ServeTheHome IPU video on YouTube, “…IPUs are going to be a big deal in the data center, and in this demo we get to show you why.”
Watch Patrick’s Big Spring Canyon IPU video titled “This Changes Networking and Storage Intel IPU Hands-on.”
Read the associated article titled “This Changes Networking Intel IPU Hands-on with Big Spring Canyon.”
Read the associated article on Intel ASIC-based IPUs titled “New Intel Mount Evans IPU ASIC DPU at Intel Vision 2022.”
You must be a registered user to add a comment. If you've already registered, sign in. Otherwise, register and sign in.