Hardware Headaches Inevitable?

JaneWalker6847 writes "Don Becker, co-founder of the Beowulf project, describes the inevitability of hardware administration headaches and warns users not to expect a silver bullet to solve these problems." From the article: "We're about to see another revolution, which is in network adapters -- that we [will] talk directly to [them] from application level. That's a massive change in how you interface with them. And that brings about a new round of device drivers completely unlike the device drivers we had 10 years ago. So, that part of the world isn't going to stabilize anytime soon."

Comments

revolution indeed

[b1ufox]

FTA

"We're about to see another revolution, which is in network adapters -- that we [will] talk directly to [them] from application level.

I hope ioctls do perhaps the same job as long as there is a module properly written to handle a specific ioctl

Or is it like controlling network firmware directly from Application ? !!!

Sounds like weird...

Imagine a malicious program kicking your Network Adapter's butt :) ...

Re:revolution indeed

[modeless]

Why would you do that when CPU cores are just about to become plentiful? I think multi-core computing should be the death of dedicated acceleration cards. When you have all these cores just lying around you don't need a dedicated sound DSP, network accelerator, or even (eventually) a video accelerator. On, say, a 64-core system, you could do real-time raytracing without any specialized hardware at all, and still dedicate a core each to network and sound if you wanted.

This kind of general-purpose architecture is far more flexible and in the end faster than dedicated accelerators.

Re:revolution indeed

[hpavc]

I think the downfall of the cards is the SSL standards, the card is tricked out for a small subset of calculations. So it can do IPSec + VLANS (in the case of that Intel server cand) and the cypto cards can do SSL for webservers and VPNs (but the implementation would be whatever the card did, not what apache could do -- though I assume apache could fall back).

Re:revolution indeed

[Short Circuit]

Transferring large files is not the primary use of high-bandwidth connections. Instead, streaming data and clusters get big bang out of it. 10Gb isn't targeted at consumers or gamers, it's targeted at businesses with lots of fresh data to push around.

Here's a short list of companies for whom 10Gb Ethernet likely comes in handy:

• Google
• Pixar
• IBM
• Amazon.com

And then there are systems on the lower end of the Top 500 supercomputers list.

Re:revolution indeed

[SchwarzeReiter]

Because dedicated cores can solve the same problem more cost efficiently than general cores. Imagine how many general cores would you need to get the capabilities of a GeForce card. And also I'm not an expert in graphics, but I think you would need a truck load of 64-bit cores to be able to do real-time raytracing.

Re:revolution indeed

[raynet]

Except that people are doing real-time raytracing already with todays computers. First real-time raytracing was running on 80486 or Pentium using somewhat limited 160 x 120 resolution, but best raytracers today run on dual CPU AMDs at VGA/SVGA resolution with 100000+ polygons at about 10-20 fps. And because raytracing can be parallised easily and gains almost linear speed up you can use new dual core CPUs to get even higher fps or resolutions. I also recall reading a paper claming that software raytracer on Opteron was faster than best NVidia GPU when using 5+ million polygons.

Re:revolution indeed

[Tower]

It is all a question of tradeoffs, and for most situations, the tradeoff of a little extra CPU against extra money on an adapter (particularly if it can increase latency) is a no-brainer. This is the same way with crypto offload.

That being said, if you are trying to scale (think a dozen gigabit cards running at high utilization) or a significant number of high-throughput IPSec/VPN clients, then the offload hardware can really show up as a big gain. Even the OTS gigabit ethernet cards these days support offload of some type - usually TCP checksum offload and some support large send offload which save quite a bit on the CPU, since these checksums are cheaper to do in HW than in software.

If you are running a layer on top with its own checksum or CRC (think iSCSI), this can use a very significant amount of CPU, or it can be handled in offload hardware and really save. Again, a multi-proc Xeon can certainly handle a full gig wire with iSCSI with all of the CRCs enabled, but it can't really handle 4 of them without some serious help.

Another issue is that the bulk of older TCP offload engines were firmware based - good path TCP/IP can be handled in ASIC logic instead. Expensive to deisgn and test, but much faster and more capable than a small processor trying to handle those kind of speeds. Great for scaling to many adapters under a single OS image - generally too expensive for a simple home / SOHO type of setup where the demands aren't that great.

As always, everything depends upon what kind of traffic you want to send, how much, and over how many interfaces.

Re:revolution indeed

[iamcadaver]

This logic does not follow through when you think of GPU's.

That's what this sounds like, giving the network card the kind of specialized bus and direct communication channels like the Graphics subsystem.

Which gives a whole new meaning to...

[patrixmyth]

Imagine a BeoWulf Cluster of these #$*&#@ drivers!

Ok, but seriously, maybe someone can answer me this. Why do we still need to construct massively parallel computing architectures at the platform level? Not saying we should toss the whole concept, but for the foreseeable future won't it make a lot more sense to stick with the Amazon model of chunking up into virtual machines? I know the FA says that this view is a mistake, but he doesn't explain why. Can anyone else?

Why?

[Anonymous Coward]

Isn't having a stack of software between the network card and your application a _good_ thing? Personally I like to leave the network configuration up to the OS and focus on developing the functionality of my apps.

Besides, what about hardware abstraction? If we're talking directly to the network adapter, isn't this taking a step back into the past (remember when you had to hand-code ASM to talk to various video/sound cards back in the early days of PC demos and games?)

nil nove sub sole

[spectrokid]

Come on, we have seen the same before with modems. First, modems did everything by themselves, then we started seeing Winmodems which pushed a lot of the work back to the main processor. Intelligent network cards will be more expensive than letting the CPU do all the work. Of course, if Moore's law becomes harder to uphold in the future, then decentralising the computing work might be the only way to make computers run faster. Until then, it will be a niche product for computers where TCP is a significant part of the CPU load (webservers). Just like 3D engines on video cards are a niche product for gaming PC's where rendering is a significant part of the CPU load. Oh and a little question: how easy will the upgrade to IPv6 be? Especially if it is not just the OS being involved? As far as I understand, Vista will have a nice soup where IPv4 and IPv6 are mixed in the same driver?

This guy is worth listening to

[vtcodger]

The article doesn't mention it, but Donald Becker is, I'm quite sure, the guy who wrote most of the Linux NIC drivers. I think that anything he says about hardware interfaces and their future is probably worth reading.

It looks to me like he's telling us that drivers are not likely to go away as an issue any time soon. Too bad, but if Becker says so, he's very likely right.

memory bandwidth

[Gr8Apes]

This comment makes me think again about AMD's acquisition of ATI. Would AMD put an ATI graphics core in the CPU package? (HTT allows for all the bandwidth the GPU could handle - no separate cache needed). Need a faster GPU? By the time you do - there'll be a faster CPU with a new GPU included, and this packaging might be less expensive than the current high end cards.

This combination would also work fine for 90% of the world's computer users, and possibly be much cheaper. Think Sempron with RAM and a miniscule motherboard with ports. The $100 laptop might drop in price.