Friction, Morning Routine, RoCe Meta Paper, AWS RNG, MRC, Slurm, Rail-Optimize, 800VDC, Phyllo, Approach

Manson AI: You need friction

AI Agent: narrow focus: goal, proof, steps

Bear Grylls’ Morning Routine: Cold (never get used to), bared foot, strength training, 30 minutes

RoCE networks for distributed AI training at scale: I have managed to read the paper ! Although in the AI word, two years is an eternity, I think it is still interesting.

1) Network Topology: backend network only for GPUs (RDMA nics), non-blocking. Frontend network: data ingestion, checkpointing, logging.

Pod = AI zone
 leaf = RTSW, DAC cables, shallow buffer
 spine = CTSW, deep buffers. fiber between leaf-spine.
SuperSpine = ATSW, oversubscribed, connect AI zones

intra-node -> nvlink
ROCE: cpu offloading, ethernet (standard)

collective communication library serves as the sw abstraction between training workloads and the NIC
                                 schedules verbs calls over QP (Queue Pairs)
 parallelism strategy determines collective: allreduce, allgather, alltoall
 choice logical topology:

------------------

2) Routing: work load. low entropy flows (few flows) -> ECMP bad (5-tuple udp: src/dst ip, src/dst port, protocol), burstiness, elephant flows
--

 RTSW uplinks 1:2 under-subscribed! -> expensive (short-term)
 1) QP scaling: use destination QP of Roce packet using the UDF capability in switch to increase entropy -> Enhanced ECMP -> short-term
 2) Central TE controller -> long-term: CP real-time topology end-to-end cluster, 
                                        flow matrix (flow bps) + CSPF (constrained SPF)
                                        write in switches dataplane
                                     DP: TE overrides default BGP routing policy in leaf. Use Exact Match table.
                           Not good with multiple link failures. Doesnt scale 
 3) Flowlet switching: try to improve 1 and 2. hw assistant schema. put packets in different ports in ECMP
     out-of-order: move packets only after 1/2 RTT
     load-aware path assignment: better than TE

------------------
                     
3) Transport: congestion management. Start with DCQCN. packet drops on ACK/NACK can cause prolonged Local ACK timeout (LAT)
--
 Tuning DCQCN not great (strict ECN -> minimize PFC (can lead to head-of-line blocking)

 200G, we stayed with relaxed ECN marking, allowing for buffer build up in the CTSW, while keeping default DCQCN settings.
 400G We proceeded without DCQCN. just PFC for flow control
 re-design collective library: two-stage copy

------------------

4) Operations:
 Change QoS priority of Clear to Send (CTS) messages. In RTSW ASIC, modify dsCP marking for ACK  messages
 Tuning VOQ in CTSW
 obeservability: OOS: out of seq.
                 Link flaps
                 Local ACK timeouts (LAT)
                 PFC watchdog: catch any long-duration PFC pause (>200ms)
                 buffer utilization RTSW
                 reachibility (pings)
                 constant latency monitoring loaded and unloaded (catch regressions)
                 base lines!!!

Perplexity: Hosting Qwen on Blackwel:

AWS RNG – Random Graph Network: The paper is totally out of my space, but the concept looks brutal. With an operations hat, how you troubleshot it? (ping, traceroute, link congestion, data flows patterns, etc)

MRC1 and MRC2 (OCI): Why we need planes (breakouts) and not just a big plane.

As SerDes speeds continue increasing, every microsecond of congestion creates much larger pressure inside the fabric. A 100G transport domain may be manageable. A 400G domain amplifies the same congestion into roughly 4x pressure. An 800G domain, and eventually a 1.6T domain, becomes much harder to coordinate.

This pressure appears as larger switch buffer requirements, larger congestion domains, harder retransmission coordination, larger cache pressure, larger synchronization storms, and harder thermal and power scaling inside ASICs.

At hyperscale, switch ASIC cache and transport coordination become fundamental scaling bottlenecks. Increasing switch buffer size is extremely difficult: high-speed SRAM is expensive, larger cache arrays consume significant power, thermal density rises quickly, die area scaling becomes inefficient, and routing complexity increases dramatically.

Splitting transport into many smaller lanes naturally reduces these pressures. Reliability improvements then emerge as a byproduct, because congestion, retransmission, and buffering become more distributed.

THE QUESTION: which breakout keeps the fabric at the shallowest practical Clos depth while keeping plane count and operations manageable? -> less hops, less switches, less latency

Slurm: I like the “Slurm vs. Kubernetes”

Slurm Workload Manager (short for Simple Linux Utility for Resource Management) has become a cornerstone of large-scale computing. Originally created in the early 2000s to support large-scale high-performance computing (HPC) environments, Slurm is now widely recognized as the de facto scheduler for HPC clusters. Today, it orchestrates jobs across thousands of servers and GPUs in some of the world’s most advanced computing environments. 

Interview Question: 512 GPU, non-blocking (full bisection) and 2xUFM! I really liked this. I think for once I understand the rail-optimize (fat-tree = leaf-spine). Just break one leaf-spine link, beautiful!!!

800VDC: Next step in electrical infra in DC space.

Phyllo by hand

Approach woman: curiosity and no performance. Practice. Be at peace with uncomfortable and akwardness. Rejection as learning

Netdev 0x19, RouteViews, Unexpected in Quantum Computing, BGP Bug, Scale-up fabrics, Network Engineer at Hyperscalers, CUDA in RISC-V, Tomahawk6, videos

NetDev 0x19: I would like to attend this one day, although is quite beyond my knowledge.

RouteViews Looking glass: link

Unexpected in Quantum Computing: A bit of a summary of the current state of quantum computing. In short, still quite to go (it seems you need 20 million qubits to break RSA-2048)

BGP Bug: link

Scale-up fabrics:

Network Engineer at Hyperscalers: Agree, but it shouldn’t just be for hyperscalers.

CUDA in RISC-V: good news, I hope it is developed, and RISC-V grows

Tomahawk6: A bit more visual

Videos:

Dr K: Masturbation – I really like this guy.

Janja Garnbret: just insane.

Jordan B Peterson + Michael Saylor: It is not about crypto, more about the life and path of Michael.

Veritasium: Quantum Mechanics – All paths possible: Very interesting video, ant the experiment is really nice

Veritasium: Electricity doesnt flow via the wires: I watched some time ago, and I wanted to recorded it here and try to watch again.

Scale Systems 2024, MS GenAI for beginners, Federer, Whisper WebGPU, Starlink TCP

Scale Systems 2024 (videos): GenAI Training: Short but interesting video. Main failures: GPU, memory and network cables 🙂 For the Network side, I liked this screenshot. Still they are able to build two 24k GPU cluster with IB and RoCE2.

MS: GenAI for beginners.

Federer: Effortless is a myth (without hard work there is nothing), It’s only a point (resilience, present), life is bigger than the court.

Whisper WebGPU: Real-time in-browser speech recognition

Free Matrix Multiplication: This looks like a big deal.

Starlink TCP: Very quick summary, control protocols with Selective Ack perform better. The ping analysis is quite good. Being able to see that each 15s you are changing satellite, is cool.

25519

Today I’ve received a copy of Serious Cryptography and jumped straight ahead to Chapter 12 talking about Elliptic Curves. I’m more or less aware of how RSA works, but was intruigued for a few years now about how criptography was able to achieve same security capabilities with more efficiency while dealing with less information.

Basically that is a different kind of beast based on the hard it is to retrieve the exponent from a discrete logarithm, instead of rely on the factorization of a number composed from two large primes as it is with RSA. As the size of the numbers are much lower with the same hardness to attack, the keys involved are also way smaller.

Elliptic Curve comes after the properties of the equation that the field of work is based, an equation of the curve of the type y^2=x^3 + ax + b where a and b are pre-cooked parameters that must be carefully chosen to avoid security risks.

In fact, there are two main curves used nowadays. One are a family of curves that the NIST approved, the most famous known as P-256, being the most commonly used in the industry, but also had some critisism because those a and b params in the equation were defined by the NSA and the generation is not completely publicly known. Therefore it could have potentially unknown pre-cooked attacks, althought based on the book, the general consensus by the experts is that there is no problem.

The second one is the curve generated by Daniel J. Bernstein that have pretty much the security of the NIST one and arguabilly a bit more of performance speed. It is called Curve25519 because the (discrete) field it works on is based on the prime number 2^255-19, having a=486662 and b=x .