5 ms·

Is it just me getting older / less smart, or did articles about products really start to sound like a jumbled mess or buzzwords lately? What is "Hardware trans

Is it just me getting older / less smart, or did articles about products really start to sound like a jumbled mess or buzzwords lately? What is "Hardware transport", what is "the ecosystem"? And then there is dozens of random products and technologies that I've never heard of... This sounds more like a humble brag, than an article trying to inform people about technologies that might actually be useful to them.

seanp2k2 3y ago
This is how you get promoted at Google.
- nottorp 3y ago
  Btw, now that someone got promoted, how dependent is this on Google's continued support? As we all know, interest at Google will now wane since you can't get promotions out of this any more.
stonogo 3y ago
"Hardware transport" is kind of a misnomer, because this is a networking protocol. It just happens to be a networking protocol that requires hardware acceleration on the NIC. We already had that in things like Infiniband and Omnipath, but Nvidia bought one and Intel rugpulled the other. Meanwhile ethernet has been approaching parity in terms of throughput, but TCP introduces unpleasant latency, so this is Google's NIH-flavored DIY on the topic. It's something of a rite of passage for a company to convince itself that building this sort of thing in-house is necessary, and that it will revolutionize high-performance computing in all the ways that previous, nearly-identical projects have not. "The ecosystem" is The Open Compute Project [1], a trade association which mostly puts together quasi-standards that provide targets so computer manufacturers can produce bleeding-edge gear with some hope that it will be interoperable. An example OCP production is the newer 21" racks that are starting to appear in datacenters. 1 - https://www.opencompute.org/ https://www.opencompute.org/
- jacquesm 3y ago
  > It's something of a rite of passage for a company to convince itself that building this sort of thing in-house is necessary, and that it will revolutionize high-performance computing in all the ways that previous, nearly-identical projects have not. The same happens with: - databases - encryption - operating systems - frameworks - programming languages I've seen this so many times by now it stopped being funny.
  - jeffbee 3y ago
    Are you suggesting that it would have been better for Google to use off-the-shelf databases etc? Because at their scale it seems clearly necessary to bring that in-house.
    - jacquesm 3y ago
      At Google's scale a lot of the ordinary limitations do not apply. Unfortunately many companies believe that because Google does it it must be good. It's cargo cult reasoning and the result is endless NIH projects. As for Google's 'Falcon' project: it smacks of NIH to me, but maybe their use cases are specific enough that none of the off-the-shelf bits were usable.
      - jeffbee 3y ago
        Which part of falcon, specifically, do you think duplicates off-the-shelf software that accomplishes the same things? Swift? Carousel? PLB?
        jacquesm 3y ago
        I think you are reading something into my comment that I did not actually write.
        jabl 3y ago
        There seems to be quite a lot of overlap between Falcon and what the Ultra Ethernet Consortium is ostensibly working on. As well as Amazon's Scalable Reliable Datagram (SRD) thing. All of them, in a way, are about addressing deficiencies in RoCEv2 for large scale latency sensitive networking that you see in HPC and DL training. But none of these are things you can buy today. Well, there's InfiniBand, but if you're wed to Ethernet..

seanp2k2 3y ago

This is how you get promoted at Google.

nottorp 3y ago
Btw, now that someone got promoted, how dependent is this on Google's continued support? As we all know, interest at Google will now wane since you can't get promotions out of this any more.

stonogo 3y ago

"Hardware transport" is kind of a misnomer, because this is a networking protocol. It just happens to be a networking protocol that requires hardware acceleration on the NIC. We already had that in things like Infiniband and Omnipath, but Nvidia bought one and Intel rugpulled the other. Meanwhile ethernet has been approaching parity in terms of throughput, but TCP introduces unpleasant latency, so this is Google's NIH-flavored DIY on the topic. It's something of a rite of passage for a company to convince itself that building this sort of thing in-house is necessary, and that it will revolutionize high-performance computing in all the ways that previous, nearly-identical projects have not. "The ecosystem" is The Open Compute Project [1], a trade association which mostly puts together quasi-standards that provide targets so computer manufacturers can produce bleeding-edge gear with some hope that it will be interoperable. An example OCP production is the newer 21" racks that are starting to appear in datacenters. 1 - https://www.opencompute.org/ https://www.opencompute.org/

jacquesm 3y ago
> It's something of a rite of passage for a company to convince itself that building this sort of thing in-house is necessary, and that it will revolutionize high-performance computing in all the ways that previous, nearly-identical projects have not. The same happens with: - databases - encryption - operating systems - frameworks - programming languages I've seen this so many times by now it stopped being funny.
- jeffbee 3y ago
  Are you suggesting that it would have been better for Google to use off-the-shelf databases etc? Because at their scale it seems clearly necessary to bring that in-house.
  - jacquesm 3y ago
    At Google's scale a lot of the ordinary limitations do not apply. Unfortunately many companies believe that because Google does it it must be good. It's cargo cult reasoning and the result is endless NIH projects. As for Google's 'Falcon' project: it smacks of NIH to me, but maybe their use cases are specific enough that none of the off-the-shelf bits were usable.
    - jeffbee 3y ago
      Which part of falcon, specifically, do you think duplicates off-the-shelf software that accomplishes the same things? Swift? Carousel? PLB?
      - jacquesm 3y ago
        I think you are reading something into my comment that I did not actually write.
      - jabl 3y ago
        There seems to be quite a lot of overlap between Falcon and what the Ultra Ethernet Consortium is ostensibly working on. As well as Amazon's Scalable Reliable Datagram (SRD) thing. All of them, in a way, are about addressing deficiencies in RoCEv2 for large scale latency sensitive networking that you see in HPC and DL training. But none of these are things you can buy today. Well, there's InfiniBand, but if you're wed to Ethernet..