The PCI Express (PCIe) protocol is becoming more commonplace and complex, making verification a bigger challenge than before. NVM Express (NVMe), which leverages PCIe to provide an interface to flash-based storage applications, only adds to the complexity. A July 25 EE Times webinar, now available on demand, provides a number of tips for (as the title says) "Shorter, Better, and Easier PCIe and NVMe Verification."
Presented by Gouqing Zhang, verification IP (VIP) CTO at Cadence, and Moshik Rubin, senior product marketing manager at Cadence, the webinar provided hands-on details about the following:
I found the discussion of performance analysis and optimization particularly interesting because, as Rubin said, this is something that is not typically done during the IC verification process. "But we think it should be, especially with NVMe coming into the game," he said. In the webinar, Rubin showed how performance can be optimized as part of the PCIe and NVMe verification process.
An Evolving Protocol
Zhang noted that the cloud revolution and the mobile revolution are placing increasing demands on SoC and system designers -- speed, cost, latency, bandwidth, low power, and resource utilization, to name a few. PCIe has addressed the speed issue by moving from 2.5G transfers/second in its first generation to 8G transfers/second in Gen3, with 16G transfers/second on tap for the emerging Gen4 spec.
Rubin identified several verification challenges. He spoke of the need to ensure that a PCIe core is functional and "bug free" before integration with a higher-level application. A second challenge is to reduce effort and to reuse much of the existing testbench for derivative protocols. A third is to set clear and measurable signoff criteria. The fourth challenge is performance optimization.
Taking IP Verification "Off the Table"
"We keep seeing many design and verification teams struggling with integration," Rubin said. "You need to integrate your PCIe design with some external model or VIP, and you need to know a lot of details of the spec and the IP." What's needed is a dedicated tool that removes the need to learn some of these details. Such a tool would have a knowledge of the protocol and would capture all relevant PCIe related parameters. It would disallow incorrect combinations. (Rubin didn't directly discuss the Cadence VIP configuration GUI, but it was displayed on one of the slides during this discussion. You can read more about it here.)
Creating an Executable Verification Plan
To cope with growing verification complexity, Rubin said that design teams should create an executable verification plan (vPlan) that is linked to the PCIe spec. The vPlan defines the needed testing for each feature, and provides and tracks coverage metrics that can be measured as verification proceeds. Once a comprehensive vPlan exists at the block level, it can be integrated into an SoC view that considers application-relevant scenarios.
Analyzing and Optimizing Performance
Functional verification checks logical bugs only, Rubin noted, "but performance bottlenecks are bugs that are sometimes even more severe than logical bugs." All too often they are exposed late in the process, perhaps in the lab, and they can lead to significant project delays or even cancellations.
The right way to handle this situation, Rubin said, is to use the same infrastructure and testbench to get performance metrics. "You have a lot of information in your VIP, or in your model, about what is going on in the link, and you can abstract the relevant metrics to help you figure out if the design is behaving in an optimal way," he said. For example, the duration of low power states can help analyze power consumption, actual data traffic can provide an estimate of throughput, and queue utilization gives insights into overall resource utilization.
PCIe provides a number of parameters that can help optimize performance. Users can adjust device-level power states, link-level low power states, the number of lanes, and number and depth of queues, and the speed of the link. "The right way to do it is to run different tests with different parameters and see what your performance is," Rubin said.
NVMe Command Processing Flow
Rather than give a generic description of NVMe, the webinar focused on the complex NVMe command processing flow. Rubin walked through this flow and talked about what needs to be verified. It adds up to a hefty list. He pointed to two ways to reduce the effort. One is to use a "strong infrastructure" that enables control of all NVMe/PCIe functionality, including error injection and corner case scenarios. Another is to create an executable verification plan and to use it to track progress.
Cadence offers PCIe and NVMe VIP, and recently added a performance measurement utility to its PCIe VIP. For further information about Cadence VIP, see the Cadence VIP Catalog. To access the on-demand EE Times webinar, click here.