[Beowulf] 10GbE Adapter Market

Lux, Jim (337C) james.p.lux at jpl.nasa.gov
Tue Nov 19 14:42:55 PST 2013


From: Bob Drzyzgula <bob at drzyzgula.org<mailto:bob at drzyzgula.org>>
Date: Tuesday, November 19, 2013 6:26 AM
To: Mark Hahn <hahn at mcmaster.ca<mailto:hahn at mcmaster.ca>>
Cc: "beowulf at beowulf.org<mailto:beowulf at beowulf.org>" <beowulf at beowulf.org<mailto:beowulf at beowulf.org>>
Subject: Re: [Beowulf] 10GbE Adapter Market

<snip>
 Then again, from another perspective it is perhaps the case that this only became possible when the quality control standards used in commercial manufacturing processes finally rose to a level that one actually *could* depend on a COTS product when the life of a soldier or astronaut depended on it.

> yes, but I refer you to MIL-P-43988C (15 June 1993), which is a specification for "Paper, Toilet Tissue"; it calls out detail such as testing according to ASTM D 3905.  Now, the stuff I buy at the supermarket may well have been subjected to such a test as a part of normal manufacturing, but, on the other hand, I'm not asking for a certificate of compliance from the vendor to that effect.  This kind of thing does drive up the cost and hassle, but, in some ways, is necessary.  Someone, somewhere must have sold an unsatisfactory product…


But from this perspective, very little computing equipment in use today is anything but COTS. Perhaps some of the giant web services companies are specifying custom computing devices, but I'd guess that the volumes involved there are high enough that distinction becomes meaningless. It is almost as if the Beowulf community taught the industry and the market something that they more or less learned.

> In the space business, very little is actually COTS at the "computer" level. Sure, connectors and even components might be. But at the "box" level (or even "board" level), production volumes are small enough that when the mfr goes to build the next batch, some of the parts don't exist any more, triggering a "requalification" of the part.  It is true that we do a lot of "build to print", but even there, you have the "part no longer available" issue.  This really hits hard when you're looking for spares.  Say your spacecraft has two flight computers, and you happen to have two of those computers sitting around (as spares from a previous mission, maybe). But you want a third unit, just in case one of the two fails during the whole assembly and test process.  So it's go off to the mfr and say " can you build us a copy of these two old computers we have?"

Perhaps the bigger remaining distinction, from an HPC perspective, is between devices built around merchant silicon vs custom or captive designs. While there is still a market for devices built around Power and SPARC processors they are certainly not the focus of HPC cluster computing (I note that less than 10% of the current Top 500 use anything but Intel or AMD CPUs; Power/PowerPC accounts for most of those). Network equipment has remained something of a bastion for products built around custom ASICs, but even that is starting to crumble as designs built around the latest merchant chips such as the Broadcom Trident II, together with protocols such as SPB & TRILL, become competitive with big-iron chassis switch configurations.

> and even in the bespoke space area, we do a lot with "commodity" CPU cores in FPGAs or ASICs.  For instance, the SPARC V8 is available as a chunk of silicon with leads, planted on a board with other parts, or as VHDL that you can instantiate in an FPGA, and you can even get it already programmed into an FPGA, with your choice of peripherals.  This is becoming a lot more popular than, say, buying a RAD750 board (which is a older PowerPC based CPU).  There are a lot of vendors selling space qualified PowerPC cards of one sort or another in a cPCI form factor, although the expected environments for those cards may be fairly benign (something going on Space Station doesn't need to handle much radiation, for instance.. People live there, so dose is low).


When I think about the "SOHO" designation, though, I am thinking more in terms of the commoditization of trailing-edge technology. Take a meander down to MicroCenter and you'll find that "SOHO" network switches are still being sold in high volumes with 100 Mbps ports; only the premium models have gigabit ports. As a practical matter, a 100 Mbps switch is perfectly suited to the task of connecting the half dozen workstations of an insurance agent's office to a couple of printers and a 25 megabit Comcast Internet link. This is the kind of application that those magazine writers were thinking of when they coined the term, and it is only incidental that some of that equipment will occasionally be useful to on-a-shoestring level HPC.


>  I agree.


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.beowulf.org/pipermail/beowulf/attachments/20131119/7e4c07a7/attachment.html>


More information about the Beowulf mailing list