Lots of posting e.g http://www.engadget.com/2013/01/16/facebook-open-compute-project-splits-up-monolithic-server-designs/ following last week's Open Compute Summit.
Silicon photonics in servers - still looks a lot like science fiction, and more importantly, for a sector which is incredibly cost conscious, the product demos on show are at least one turn of the Si crank away from a fully integrated photonic backplane interconnect.
This integration step needs (maybe finally after about 15 years of the industry flopping around and reinventing the wheel) to address the fundamental issue of whether the transport in the rack/chassis should look like a network (aka Ethernet) or a bus (aka PCIe). The obvious answer is of course that there are 'horses for courses' and it should be both.
Micro-web servers (and software endpoints in general) want to talk to something that looks like an Ethernet network and bits of hardware want to talk to each other using something that looks and feels like an IO bus.
Yes there are exceptions - example: use cases for shared IO devices which would allow software to talk to a network controller yet naturally present as a harwdare (bus centric) endpoint over the interconnect. But it is always worth bearing in mind that contrary to popular belief, the chip cost for the bus centric interconnect to a shared IO device is actually very little different from the silicon cost of a non-shared device.
It's worth looking back at history the Infiniband roadmaps at the turn of the Millenium defined a new system model for both software and hardware to adopt and at the same time major chip manufacturers and OEMs were publicly stating that this was to be the future of server architecture south of the 'NorthBridge' . Instead bus architecture evolved to PCIe, networking has remained Ethernet, SANs have stayed FC and are evolving into iSCSI and decentralised architectures driven by flash storage.
The bow wave of the IB investment has dispated over a decade, leaving behind one IPO and a high-end HPC interconnect, whose value is simply that of staying one node ahead of Ethernet on the bandwidth curve.
Maybe it was one long hangover after the party, but in terms of wasted R&D dollars, the last decade has been a big one for major wastage and generally poor results. A good over-a-beer conversation would be to total up the $ cost of:
1. TCP Offload Engines - now pointless given multicore CPUs
2. FibreChannel over Ethernet - going no where - a great cover to enter the server business
3. RoCE - going no where - value of IB is the physical layer - nothing more
4. DataCenterEthernet - only useful for layer violating protocols like FCoE and RoCE. Also many of the standards require significant interop testing by competing Ethernet switch vendors (spot the oxymoron)
5. All those dumb protocols (that OEMs spent a long time assiduously courting) iSER, iWARP, SDP ...
Which takes me back to the Open Compute Platform. Mega data center operators have long held the view, and now large main street customers have decided that OEM driven technology initiatives have failed - that the current value add is largely the metal box.
The result is that in order to command product margin, true value add needs to be established in the customer base in the form of deployable product with ROI. This is the matra that Solarflare operates to.
The is the new mantra for everyone who wants to stay in business.
Happy New Year! -- oh BTW I'm tweeting on slp00000 - likey to continue for a while to win a bet.
