ANTONY SAVVAS TEST & MEASUREMENT
OPTIMISATION For optimising AI networks through testing, Ram Periakaruppan, VP and GM, network test and security solutions at Keysight, says you have to be strategic. “To properly optimise AI networks, no single product will do everything. At one end of the development cycle, before a new AI accelerator chip is even cast into silicon, test solutions must integrate with chip emulation environments and recreate the large AI networks in which this new chip will operate,” Ram says. “At the other end of the cycle, you have to rely on simulation environments to assess the performance of tens of thousands of GPUs that collectively work to train an AI model. In between, lies a huge set of test cases where it’s critical to execute highly accurate, hardware- based emulations of AI racks to carefully calibrate system performance.” Tying these together is a shared framework that defines an AI workload and the resulting network traffic, and which ensures test cases are “portable” throughout AI system development, deployment, and operation, adds Ram. DEEPER INSIGHTS So what are companies now testing for? Testers in the AI ecosystem are looking for deeper insights that connect system- level aggregates with performance of low-level components in the AI cluster. For example, says R am , the job completion time of a training batch has a tighter integration with business fundamentals, and it almost directly translates into cost. “During that batch, there are multiple collective data exchanges over the network. So we must drill down into each collective completion time, which is dependent on congestion control metrics, load balance efficacy, and ensuring no packet loss across the fabric,” he explains. Some AI models are known to be network bound, but you can’t simply optimise networks and necessarily
framework and AI governance, and establish a link between AI/ML, open data architecture and intent-based autonomous networks.” While the promise of AI is “impressive and real”, there is no “silver bullet, single answer” for the test and measurement sector, but the overarching trend is clear, Farid says. “AI can help test and measurement providers move from simply gathering and providing data, to more effectively and efficiently solving the biggest connectivity issues for the greatest number of users.” Yamany from VIAVI confirms, “Moving forward, expect the T&M ecosystem
expect them to perform better. Measuring job completion time means the test device must emulate the full workload since each AI model has its own unique way of distributing work and coordinating results. “Not only is the measurement changing, but what the test system must create is also changing,” Ram says. CHANGING PRODUCT SET? So with a wider take-up of AI, is their a greatly changing product set in the testing arena? Yokogawa Test & Measurement told Optical Connections: “We’re addressing AI’s impact on optical networks by developing advanced tools and solutions. But currently, AI isn’t drastically changing our product
to become more integrated with AI and machine learning, leading to more proactive and predictive network management solutions. Future developments will likely
sets. We’re closely monitoring AI developments to integrate useful
focus on automated, scalable testing solutions that can adapt to the rapid advancements in optical networking technologies, fostering a new age of efficiency and innovation in high-speed data transmission and connectivity.” “The T&M ecosystem for Ethernet must expand beyond just networking, because AI is a systems problem,” says Ram from Keysight. “You can perform basic functional network component testing that passes in isolation yet fails when the same component is deployed in a realistic AI system. To properly recreate what that system looks like, test equipment requires a workload perspective, otherwise the tester simply won’t stress the component in a meaningful way. This systems approach is the largest departure for the T&M ecosystem, along with the idea that test products span the complete development cycle, from pre-silicon to deployment, and yet share test cases and data models that apply throughout.” Like in most industries and sectors, the advent of AI is creating a challenge for the optical test and measurement industry, but it does seem up for that challenge.
advancements when ready. And we’re exploring how AI can improve predictive maintenance and data analysis in our products, gradually incorporating AI without major disruptions.” Aniket Khosla, VP of Wireline Product Management at Spirent, said of new products, “Test and measurement companies are introducing new solutions to test the capability of the AI infrastructure. We are adding products to test the efficiency of Collective Communications Library algorithms in transmitting data, for instance.”
T&M EVOLUTION As for the test and measurement
ecosystem changing going forward, Ookla’s Farid says, “Over the last 12 months, we’ve seen AIOps start to surface in the industry with the objective of automating processes across the organisation and proactively addressing issues before they materialise.” This covers the likes of network capacity, root cause analysis, and quality of experience. We’ve also observed a strong push from the TM Forum to evolve existing models and operations to include AI, such as an AI/ML
(Image: HUBER+SUHNER)
Hamdy Farid Senior Vice President of Product at Ookla.
Sameh Yamany Chief Technology Officer at VIAVI.
Ram Periakaruppan VP and GM, Network Test and Security Solutions at Keysight.
Aniket Khosla VP of Wireline Product Management at Spirent.
22
www.opticalconnectionsnews.com
| ISSUE 38 | Q3 2024
Made with FlippingBook Digital Proposal Creator