Amrit Kumar

Pls join our session @ NetApp Insight #NetAppInsight 1392-2 NetApp Data Backup — all up! By Rohit Kapoor Ebin Varghese Kadavy Jacob John Andathethu #NetApp

97

4 Comments

IB Tests closely mimics real-world RDMA and communication workloads. The tests leverage InfiniBand's high bandwidth, low latency, and direct memory access features to provide insights into network performance. InfiniBand Packet Flow ---------------------- In InfiniBand, communication occurs between two endpoints: Queue Pairs (QPs), which consist of a Send Queue and a Receive Queue. The basic packet flow involves the following steps: a. Initialization and Connection The server initializes its QPs and waits for incoming connections. The client initializes its QPs and establishes a connection with the server using Local Identifiers (LIDs) or Global Identifiers (GIDs). b. Work Request (WR) Posting The client posts a Send Work Request (WR) to its Send Queue. The server posts a Receive Work Request (WR) to its Receive Queue. c. Packet Transmission The client’s Host Channel Adapter (HCA) processes the WR, creating one or more packets for transmission. Packets are transmitted through the InfiniBand fabric, adhering to the configured transport type (RC, UC, or UD). d. Packet Arrival and Completion The server's HCA receives the packets and places the data in the designated buffer. A Completion Queue Entry (CQE) is generated in the server's Completion Queue (CQ), indicating the operation's success. e. Acknowledgment (if required) For reliable transport types like RC, the server sends an acknowledgment back to the client, completing the transaction. Message Exchange in PerfTest ------------------------------- PerfTest tools measure network performance by exchanging messages between client and server. Here's how messages are exchanged: a. Latency Tests (e.g., ib_write_lat): Measure the round-trip time for a single message. Flow: 1. The client sends a message to the server. 2. The server immediately responds with the same message. 3. The client measures the time taken for the round trip. b. Bandwidth Tests (e.g., ib_write_bw): Measure the maximum amount of data that can be transmitted per second. Smaller messages typically have lower latency but lower bandwidth, while larger messages maximize bandwidth. Larger MTU sizes (e.g., 4096 bytes) reduce packet overhead and improve efficiency. Flow: 1. The client sends a stream of messages to the server without waiting for acknowledgments. 2. The server receives and processes the messages. 3. The test duration or number of messages determines the bandwidth. c. Message Rate Tests (e.g., ib_send_bw): Measure how many messages can be sent per second. Flow: 1. The client sends fixed-size messages to the server. 2. The test reports the rate at which the messages are transmitted and received. #RDMA #DATACENTER #AI #HPC

26

Did you know that AppNeta and DX NetOps by Broadcom was recently named the highest-scoring leader and outperformer in the 2024 GigaOm Radar Report for Network Observability for the fourth year in a row? 🔥🔥🔥 Read the report to find out what makes Broadcom network monitoring solutions so superior!

7

Raj Yavatkar, CTO of Juniper Networks, candidly shares this story of his time at Intel Corporation and how he learnt to evaluate an idea from the lens of business and finance: "I was very proud of the fact I had come up with something on chip design, that will save the number of people used for testing validation. It will cost a little bit of area, but it will save about 40, 50 people. And this design manager explained to me, Raj, even if you save 100 people, he said I don't care. I was shocked why would a design manager not care? He said I make 100 to 200 million chips every year. If you save me even a half a millimeter square, it's a lot of money for me. People cost is fixed, I don't care. And until then I was so proud of my idea. It completely changed my perspective about how to think about the business, how to think about the finance, how to think about what matters to the product manager." The full episode has more stories and takeaways from Raj's journey at Intel Corporation, VMware, Google Cloud and Juniper Networks: https://lnkd.in/g6zjqeqT

3

Good points for thought. Given that GPUs are 80% of the AI cluster cost, 20% utilization leads to huge waste of power. ROI on a AI cluster may be tough for most unless there clear use cases that are connected to value delivery and monetization. It is a good point on the checklist though

6

Honnappa Nagarahalli (Arm)'s DPDK Summit NA talk breaks down packet data flow challenges in chiplet-based SoCs, focusing on latency issues and optimization opportunities for packet processing. Watch his talk and find all DPDK Summit NA sessions on our YouTube channel: https://lnkd.in/g-j2UsB3 Highlights: 🎯 Latency in chiplet SoCs due to separate compute and I/O dies 🎯 Complexity of cache snoops and DMA transfers across dies 🎯 Cache stashing as a potential optimization in future designs #dpdksummit #networking #opensource

6

Discover how integrating Palo Alto User ID-Based approach with Trisul NetFlow Analyzer empowers network visibility. Our latest blog highlights: ♦ What User ID-based analysis is and it’s importance. ♦ The key benefits of USER ID based analysis for your organization’s network ♦ How Palo Alto Networks integrates with Trisul NetFlow Analyzer 👉 Read more: https://lnkd.in/gDscWZP3 🔔 Stay tuned for the latest updates! #trisul #netFlowanalyzer #networkvisibility #latestblog #paloalto #networks #networkmonitoring #unleashnetworks

5

🚀 Broadcom joined the $1 trillion market cap club silently last week. 🔍 𝗪𝗵𝗮𝘁 𝗗𝗼𝗲𝘀 𝗕𝗿𝗼𝗮𝗱𝗰𝗼𝗺 𝗗𝗼? Like NVIDIA, Broadcom also is known for making AI chips for tech companies like Google, Meta, Bytedance, and more. They also recently signed a deal with Apple 📈 𝗪𝗵𝗮𝘁’𝘀 𝗗𝗿𝗶𝘃𝗶𝗻𝗴 𝗜𝘁𝘀 𝗚𝗿𝗼𝘄𝘁𝗵 𝗶𝗻 𝗙𝗬24? 💻 Software Sales: Nearly tripled to $5.8B 🔧 Semiconductor Sales: Rose by 12% to $8.2B 🤖 AI Revenue: Exploded by 220% last year to $12.2B 💡 𝗪𝗵𝗮𝘁’𝘀 𝗧𝗵𝗲 𝗕𝗶𝗴 𝗗𝗲𝗮𝗹? Big tech is set to invest nearly $220 billion in 2025 on AI infrastructure, with leaders like Meta, Microsoft, Amazon, and Alphabet Inc. at the forefront. 🏢 𝗕𝗿𝗼𝗮𝗱𝗰𝗼𝗺'𝘀 𝗛𝗶𝘀𝘁𝗼𝗿𝘆 🔄 Origins: Began as Avago Technologies, spun out from HP’s semiconductor division. 💼 Major Acquisitions: LSI, an Avago Technologies Company, Brocade, CA Technologies and the landmark VMware acquisition.

4

Premalakshmi R, in a conversation with TechCircleIndia, highlights how the convergence of AI and multi-cloud will redefine businesses in 2025 https://lnkd.in/gpQgt8up

2

To call Cisco's recent restructuring "overdue" is a big understatement. Look no further than the company's rapid market share erosion in the telco vertical over the last few years. Just as Cisco was expecting a recovery as telcos started revamping their IP backbones and moving to stand-alone 5G, the vendor's telco market share has plunged: from 4.2% in 1Q22, it ended 1Q24 just below 1%. This kind of rapid loss of market share is rare in the telco sector, where buying decisions are usually long-term commitments. It also came as a time when one major rival, Huawei, has been hobbled by legal restrictions. Fierce Network story: https://lnkd.in/g9WTc_uz MTN Consulting's latest telco network infra market share report: https://lnkd.in/g2ftHa85 Cisco's market share in telco NI, per MTN Consulting, is shown in the attached visual. #telco #telecom #vendors #marketshare #layoffs #restructuring

2

Now we are talking #vSAN ! Rakesh Radhakrishnan is talking about the future of #vSAN and that we expect that #vSAN and traditional storage will co-exists in the future. I can tell you great futures in the DR and Replication space. Lots of great discussions and request. VMware by Broadcom

62

As enterprises build their edge strategies, they're looking to telcos for guidance on distributed networks. 📡 Watch Broadcom experts discuss the role of telcos in enabling the future of the edge! https://brcm.tech/4e9JTZU

1

Remember the days of juggling OS, network, database, and system engineers to keep your infrastructure running smoothly? Those were the days! 🤯 It used to be a complex symphony of coordination, with OS engineers fine-tuning CPU usage and RAM, network engineers ensuring ironclad connectivity, DBAs managing intricate schema updates, and system engineers and architects tackling a myriad of other challenges. It was a juggling act that required constant attention and expertise. Cloud computing has revolutionized IT operations, condensing these roles into one powerful position: the Cloud Architect. Cloud Architects are the modern-day IT superheroes, possessing a deep understanding of everything from infrastructure to applications. They design, implement, and manage cloud solutions, ensuring optimal performance, security, and cost-efficiency. From architecting scalable systems and optimizing resource utilization to implementing robust security measures and ensuring high availability, Cloud Architects are the linchpin of successful cloud adoption. If you're looking to streamline operations, reduce costs, and unlock the full potential of your business, a Cloud Architect is your secret weapon. Let's discuss how cloud computing can transform your organization. #cloudarchitect #digitaltransformation #cloudmigration #infrastructure #integriti

5

It is indeed getting harder to keep up with the innovations in AI cluster fabric designs 🤓... Alibaba's paper at #SIGCOMM24 is another good read. Their 15K GPU High-Performance Network (HPN) focuses a lot on having a fault tolerance design with seamless failover while optimizing network performance. A few highlights💡 ✔ They note that ToR switch failures could significantly impact training performance. They mitigate this by designing a dual-ToR system - each of the 8 NICs in each server is connected to two different ToR switches in an active-active configuration where both ToRs handle traffic actively. If one ToR fails, the traffic of the failed ToR is automatically and seamlessly rerouted to the remaining ToR without interrupting the training.   With a rail-optimized topology, even without a dual-ToR design, one could theoretically reroute traffic through other NICs inside a server when one rail switch fails. However, they didn’t adopt that design point, as it needs significant modifications to NCCL.   ✔ The architecture uses a dual-plane design, where ToR switches are divided into two groups (planes), each connecting to a different set of aggregation switches. This helps to reduce hash polarization and ensures that traffic is balanced across the planes. Traffic from one plane is isolated from the other, which prevents congestion from spreading across planes. If there is a failure in one plane, the other plane can still function, maintaining network resilience. ✔ They discuss the vapor chamber heat sink design for the 51.2Tbps switch chips. The heat sink has more wicked pillars in the center to align with the hot spot in the center of the chip. This points to the thermal challenges of the densely packed chips and how thermal design is becoming increasingly complex.   ✔ For congestion control, they use a bit fancy scheme to get around the low entropy problem with ECMP hashes. The host initially establishes multiple disjoint (not going through the same set of switches) connections for each new GPU-to-GPU connection. They implement a load-balancing scheme, where, for each connection, the host counts the total number of bytes currently in the Work Queue Elements (WQEs) of all the disjoint paths. If a connection becomes congested (larger count), it will direct new messages to an alternate connection with the smallest count, i.e., the least congested path.   It is interesting to see how each hyperscaler is adopting different (often complex) congestion control strategies to dynamically route traffic through the least congested paths in the network... With UEC-compliant switches/NICs in the near future (wishful thinking), congestion control should get a lot simpler IMO - driven by sender/receiver-based congestion control mechanisms and a new transport protocol that allows unordered delivery, enabling switches to just spray evenly at the packet level across least congested links... Any thoughts? 🤔 #LLMs #genAI #AI #junipernetworks #SIGCOMM

264

28 Comments

Did you catch our Application Delivery Top Ten List? No? Well, let me bring you up to speed then. It’s all right here—the why and the what - laid out for you in one convenient spot. Trust me, this list will be a game-changer as you start designing #AI infrastructures to support the wide variety of AI applications you'll be relying on in the future. And here’s the thing: it’s not just helpful—it’s crucial. What’s happening in application delivery right now is nothing short of revolutionary. Get ready for what’s coming. 🚀 #performance #appdelivery #security #adc https://lnkd.in/gqeDXs7U

33

"Thrilled to have arrived at ElasticOn Bangalore! The energy here is contagious. Excited to spend the day engaging with fellow developers, architects, IT pros, and SOC analysts as we explore the forefront of #Search, #Observability, and #Security. Can’t wait to see how #GenerativeAI is unlocking new potential through the Elastic Search AI Platform. Looking forward to gaining valuable insights from industry experts and thought leaders. Let’s seize this opportunity to innovate and connect! Elastic #ElasticON"

52