Love is not enough

Before moving, I went through some piles of papers and found a couple of pieces with the note “Love is not enough”. I am pretty sure they were notes I took from a book of Mark Manson (likely The Subtle Art of Not Giving a F*ck – I need to re-read) And somehow this morning, I got this video in youtube.


Love is not enough:

  • Love doesnt equal compatibility: Love is emotional. Compatibility is logical
  • Love doesnt solve your relationship problems (amen)
  • Love is not always worth sacrificing yourself: Dont tolerate behaviours in your partner that you wouldn’t do with your best friend. A loving relationship is supposed to supplement your individual identity, not damage it or repace it.

The only way you can fully enjoy the love in your life is to choose to make something else more important than love.

Friendship > Love

In my case, I dont really think about love but companionship. I think too many times about that “perfect” companion to go through life (good and bad moments). And then I am very solitary person (it gets worse with age). And then you are not sociable, and many things you read tell you that having a strong social support is critical in mental health, life satisfaction, etc. So it is like in the middle of two paths.

Good concert of NIN.

Networking Scale 2023

This is a conference about networks that I was interested and I finally got some emails with the presentations. They are mainly from Meta.

Meta’s Network Journey to Enable AI: video – second part interesting.

  • AI fabric (backend: gpu to gpu) hanging from DC fabric.
  • SPC (Space, Power, Cooling)
  • Fiber, Automation
  • RDMA requires (lossless, low-latency, in-order) -> ROCEv2(Ethernet) or IB
  • Servers have 8x400G to TOR. Tor 400G to Spines
  • 1xAI zone per DH. 1xDC has several DHs.
  • Oversubscribed between zones, eBGP, ECMP.

Scaling RoCE Networks for AI Training: video — Really really good.

  • RMDA/IB used for long time in Research.
  • Training: learning a new capability from existing data (focus of the video)
  • Inference: Applying this capability to new data (real time)
  • Distributed training for complex models. GPU to GPU sync -> High BW and low/predictable latency.
  • ROCEv2 with (tuned) PFC/ECN. TE + ECMP (flow multplexing)
  • Oversubscription is fine in spine (higher layer)
  • Challenges: Load balancing (elefant flows), Slow receivers/back pressure, packet loss from L1 issues (those flapping links, faulty optics, cables, etc xD), debugging (find jobs failures)

Traffic Engineering for AI Training Networks : video – interesting both parts.

  • Non-blocking. RTSW=TOR. CTSW=Spine. Fat-Tree Architecture. 2xServer per rack. 1xserver=8xGPU. CTSW=16 downlinks -> 16 uplinks. Up to 208 racks?
  • ROCE since 2020. CTSW are high redix and deep buffer switches.
  • AI Workload challenges: low entropy (flow repetitive, predictable), bursty, high intensity elephant flows.
  • SW based TE: dynamic routing adapted on real-time. Adaptive job placement. Controller (stateless)
  • Data plane: Overlay (features from broadcom chips) and Underly (BGP)
  • Flow granularity: nic to host flow.
  • Handle network failures with minimum convergence time. Backdoor channel with inhouse protocol.
  • Simulation platform. NCCL benchmark.

Networking for GenAI Training and Inference Clusters: video Super Good!

  • Recommendation Model: training 100GFlops/interation. inference: few GFlops/s for 100ms latency.
  • LLM: training 1PetaFlops/sentence (3 orders magnitude > recommendation), inference: 10PF/s for 1sec time-to-first token. +10k GPUs for training. Distributed inferencce. Need Compute too.
  • LLama2 70Billion tokens -> 1.7M hours of GPU. IB 200G per GPU, 51.2 TB/s bisection bw. 800 ZetaFlops. 2 Trillion dataset. 2k A100 GPUs. As well, used ROCEv2 (LLama2 34B).
  • +30 ExaFlops (30% of H100 GPUs fp8 peak) + LLama65B training < 1day.
  • Massive cluster: 32k GPUs! Model Parallelism.
  • LLM inference: dual-edge problem. Prefill large messages (High BW) + Decode small messages (latency sensitive).
  • Scale out (-bw, large domain. Scalable RDMA (IB or Ethernet), data parallel traffic) + Scale up (+BW, smaller domain. NVLink 400G, model parallel traffic)
  • 32k GPU. TOR (252), Spine (18), AGG (18). 3 levels. Oversubscription Spine-Agg 7:1. 8 clusters. 252 racks per cluster. 16 GPUs per rack (8x252x16=32k GPUs). ROCEv2!
  • Model Parallelism harder for computation. Model Parallel traffic: all-reduced/all-to-all, big messages (inside cluster = sclae-up). Data Parallel traffic: all-gather & reduce-scatter (between cluster = scale-out, NVLink)
  • Challenges: Latency matters more than ranking. Reliability !!!!!
  • LLM inference needs a fabric.

Scale out vs scale up: storage DB

scale up (vertical): more bw(links), more storage, etc

scale out (horizontal): distribute load into different devices

Network Observability for AI/HPC Training Workflows: video

  • ROCET: Automating RDMA metric collection and analysis for GPU training. Info from hosts/nics and switches.
  • Report: out-of-sequence, nic flaps, local ack timeouts.
  • PARAM + pytorch. Chakra.


From ESNOG (although I am subscribed, I dont really receive notifications…) I saw this presentation, and then I found the main video reference.

BUM: 1.5Mbps -> that goes to all customer ports!! Small devices can’t cope with that.

Ratelimiting for BUM ingress/egress

Originally IXP is VPLS. Issue with long lasting flows / long lasting MACs

Solution: EVPN + ProxyARP/ND Agent.

I know EVPN is mainly for Datacenters but it is interesting the move to IXP. Although I remember reading about EVPN into Cisco IOS-XR platform that is mainly a ISP device.


I didnt know anything about this conference until the last two months started to read news about it from different blogs. I am surprised the webpage doesnt link to the videos. This one is quite interesting but after the 15 minute becomes very hardcode for me.


I guess there are hundreds of project that try to use GPT to build apps. I found this one and looks very nice. I would like to use it if I find time for one idea.


From this blog, I could read an interesting presentation about network performance using satellite services. I guess there are many blogs about the performance of Starlink but this was the first time I read something about it. I was surprised the results were not that good even with the latest version of Startlink (that I didnt know neither)


I checked out this blog note from Google about Falcon. To be honest, I dont really understand the “implementation”. Is it purely software? Does it interact with merchant Ethernet silicon ASICs? There is so much happening trying to get Ethernet similar to Infiniband that I am wonder how this fits outside Google infra. At the end of the day, Ethernet has been successful because everybody could use it for nearly anything. Just my opinion.


I watched this video yesterday by coincidence . I was keen to know about making learning addictive. And honestly, I liked the video, it was funny and informative. And it is interesting that you can turn evil manipulations into positive ones. I didnt know the presenter was the founder of duolingo and he was from Guatemala. I use duolingo to learn languages, I am not sure if you can really learn a language this way (at least I am not that smart) but it helps to practice a little, and then the compound interest should get you to some decent level. This remind me to brilliant. It is something I wanted to use, but at the end of the day, I dont have time for all things I would like to do. I need to focus in the things I really need to do.