Hacker News Reader

Comment by nickpsecurity | Hacker News Reader

Back when I looked into high-speed networking in small spaces, my idea was doing it over PCI directly bypassing Ethernet. I really wanted to do something like SGI's NUMAlink but PCI be cheaper with commodity parts available. You'd need a switch for it, though. I ended up finding a company selling them.

So, there's always that kind of thing to consider. I wonder if 10-100Gbps have better price/performance than TCP/IP over PCI by now. I cant even rember the company or tech's name.

View on HN

I wanted to do the same thing with PCI!! I thought the bandwidth was large enough that it should be a great direct interface. Unfortunately synchronizing between two pci masters was going to be tricky and the high frequency development required for PCIe requires some pricey diagnostic equipment. I ran across those off the shelf switches / interfaces, but they were a lot more expensive that 10gb ethernet at the time and they weren't even available for high lane count pcie 2 and pcie 3.

I think now that 10gb Ethernet prices are coming down to about $100 per card and switch port there's no way specialized direct pcie would compete off the shelf.

Still I think it would be cool to have a direct 32 lane pcie 3 of 4. It could operate as fast as ram speeds. If you connected two 2P beasts it would be close to having a 4P.

Of course there would probably be protocol issues and overhead as the no-two-masters issue would mean you can't use the native protocol directly.

I didnt think about the specialized, diagnostic gear. That couldve priced me out.

Far as your other idea, a Distributed, Shared Memory setup was one of my potential applications. I was looking at Beowulf clusters and old MPP systems trying to find ways to get more CPU's with large RAM and simplified programming. I found quite a few software implementations of DSM model. Infiniband and Myrinet were main thing in networking. So, PCI idea was partly about connecting them with something that already had DMA. That at least led me to NUMAscale's products for scaling AMD systems.

Note this comment in this subthread: https://news.ycombinator.com/item?id=17276861

Don't worry, the chip you are searching is branded ExpressFabric and sold/made by Broadcom. It's AFAIK only PCIe4, and you get ones with up to iirc 96 lanes per chip. They support to be connected into a mesh, even adding shared PCIe devices like network cards, GPUs, NVMe and SAS/SATA HBAs is possible, though you might have difficulties with the GPUs in particular. The others, if you are careful when buying, should do well if they can handle SR-IOV, as (from their view), the mesh switch they are attached to is the root, and the computers attached to the mesh take the place a VM would normally occupy.

If you could get the smallest chip they make, it should not be hard to get a PCB made for it that fans out to e.g. a set of USB-C receptacles. One such receptacle can handle 2 PCIe3 lanes, and some power/USB-2 on the side. The main reason being the relatively low price, compared to e.g. mini SAS HD, which has a similar density and twice the lanes per connector, but also a slightly higher target impedance at the upper end of the PCIe spec, whereas USB-C is specifically targeting the perfect PCIe impedance.

A quick search on Digikey shows several entries for ExpressFabric which seem like what you mean:

https://www.digikey.com/products/en?keywords=expressfabric

They all have a part status of "Discontinued at Digi-Key" though. Not seeing any further info of where they _are_ available either. :/

You found the chips I was referring to. If you want them, you should probably ask Broadcom, as they should either be able to point you to where you can still buy them, or to a replacement. And in case they don't help you, you can still ask sales at competitors, as some might have something suitable. This certainly looks like they'd do a run for you if you want some, albeit with a more significant MOQ (probably in increments of one wafer, with the best price to you if you agree to take all good dies of those they make for you). According to their website [0], they are still active products. It just seems that no one really wants to buy them, due them being somewhat weird. There is a reference platform [1], which is a 32-port, as far as I can tell PCIe3 x4 on QSFP ports, 1U TOR switch. It goes with 2-port cards that go into the servers and contain re-timers. Broadcom claims one can connect the switch with the server cards using optics or copper, but the switch is $11k at mouser. The technology apparently offers virtual Ethernet NICs and 8 QOS classes.

Did you want to buy such technology?

[0]: https://www.broadcom.com/products/pcie-switches-bridges/expr... [1]: https://www.broadcom.com/applications/datacenter-networking/... (bottom of the page, grep "ExpressFabric Reference Platform")

> Did you want to buy such technology?

Nope, just merely curious

For most of the use cases I can think of that these would suit, Infiniband seems like it would also work and that's fairly widely available already.

That being said, there's probably use cases this would suit better. They're just not coming to mind easily. :)

Cheaper than Infiniband, I assume. Even if you buy the TOR switch, from what I can tell.

Also, the latency is much smaller than for Infiniband, and it is nice to e.g. combine blades with NICs or similar configurations, without the NICs being special multi-host ones like Mellanox offers.

That's possible with PLX/Broadcom ExpressFabric but those switches have fallen far behind Ethernet. The fastest PCIe switch is ~768 Gbps while ~6500 Gbps Ethernet switches are available.

You need to consider that these PCIe switches you talk about are (afaik, at least there are big ones) single-die, compared to the probably even multi-board construction of the Ethernet switches. They do support and are in at least one document advertised to be used in a switching fabric, where you create a mesh between the nodes you try to connect. Basically, instead of plugging every computer into the same PCIe switch, you fill the PCIe switches half full with computers, maybe even a little less/add some that don't connect directly to computers, and add some interconnections between the switches. The mathematics that describe how one would want to arrange the small switches date back at least to when automated telephone switches became common. A similar version, though with separated inputs and outputs, is known as a Crossbar Switch.

Did you by chance come across something smaller than a full-blown, PCIe3x4 per node, TOR switch? I don't need that much capacity, but a cheap enough high speed switch would allow using some slower (individually) servers that got kicked out from the previous owner, instead of getting new, larger machines. A certain level of microservice-thinking/scale of (distributed) Haskell/Erlang does often require much faster connections than one would like to afford with Ethernet, to keep the coding overhead for the sharding to a minimum. At a certain point one can stop worrying about minimizing transfers, and just rely on page-faulting and similar things for less-critical (in a performance sense) data accesses.

I was talking about single-chip switches. The thing about Clos is that it has non-linear scaling, so if you use a chip with lower capacity you end up needing far more chips.

All of these chips have smaller sizes available.

There was a product in late nineties that did P2P links over ATA66/100 IDE interface. No drivers needed, just write/read to a virtual filesystem on simulated drive.

Reply on news.ycombinator.com