Original URL: http://www.theregister.co.uk/2006/06/06/sun_labs_mpk/

Sun Labs edges toward practicality

Games, Solaris and DARPA - Oh my!

By Ashlee Vance

Posted in Servers, 6th June 2006 21:48 GMT

You never know what will turn up when Sun Microsystems opens the doors to its Labs operations, as it did for reporters last week. In the past, the company has tried to woo observers with things as trite as drying paint - or at least drying paint full of sensors. This time around, Sun offered some more gripping - and reality tethered - labs work to the press.

Massively parallel bass fishing

Dashing to the eye-candy first, we bring word of "Project Darkstar." This is Sun's move to build video game servers, which could one day power a video game hosting service.

Sun reckons that current data centers powering on-line games are too expensive and inefficient. Many of the systems used to run Second Life, City of Heroes, EverQuest and other games falter due to heavy user volume, server failures and data errors. Sun hopes to apply its data center expertise to create a game server that has better load balancing, fail over traits and general intelligence.

Backers of today's most popular games must spend millions of dollars per month on hardware and data center operation costs. Such expenditures make life tough on smaller game developers that want to own a piece of the on-line action. By freeing up its game servers in a hosted model, Sun could tap this small developer market.

"There might be 500 people that want to play online bass fishing," said Jeff Kesselman, the lead behind Sun's game server project. "It's just not affordable to build a system that can do that today. With our system, small developers can get in."

Sun is currently in discussions with "one of the biggest developers in the business" about moving to a type of hosting model. Sun would either handle the servers in-house or team with a telco partner. Sun, however, doesn't plan to use its $1 per CPU hour grip model for games.

The games group has been testing out their software on Sun's upcoming Opteron-based blade servers code-named Andromeda.

"Technically, we're probably a year within being ready (for large-scale hosting," Kesselman said.

There's more on the project here. There's a good piece of game scaling woes here as well.

NICed and sliced

Closer to the heart of Solaris geeks is the "Crossbow" project to virtualize NICs.

Sun researchers have moved to tackle what they see as one of the next major pieces of the virtualization puzzle. Customers today run two, three and four operating systems (and related applications) per server with packages such as VMware and Xen. Those using Solaris Zones can run even more OS instances per server due to the arguably more efficient zone model. Controlling how networking loads from these partitions are handled by NICs can be difficult.

Crossbow tries to solve some of the networking problems by making sure each applications gets a set amount of bandwidth.

"Crossbow provides the building blocks for network virtualization and resource control by virtualizing the stack and NIC around any service (HTTP, HTTPS, FTP, NFS, etc.), protocol or Virtual machine," Sun says.

"Each virtual stack can be assigned its own priority and bandwidth on a shared NIC without causing any performance degradation. The architecture dynamically manages priority and bandwidth resources, and can provide better defense against denial-of-service attacks directed at a particular service or virtual machine by isolating the impact just to that entity. The virtual stacks are separated by means of [a ] hardware classification engine such that traffic for one stack does not impact other virtual stacks."

Crossbow has been set up as an OpenSolaris project, and there's more detailed information available here.

PetaSun

Sun is competing against Cray and IBM for one of the most lucrative government contracts in recent memory, from DARPA. The vendors have been tasked with creating systems that will allow the US to skip ahead a supercomputing generation with systems in 2010 that are 10x to 40x faster than today's. It looks like DARPA will hand out the deal to two of the three vendors and announce the winners next month.

Sun's design, as far we know, centers around the second version of its upcoming Rock processor. The first iteration of Rock - due out in 2008 - should have 16 cores, and the 2010 version may well have 16 cores too. All told, Sun expects to handle more than a million software threads over tens of thousands of cores with its DARPA box.

Much of the technology Sun has crammed into the DARPA bid is mentioned here.

There are still some large gaps in the plan that Sun won't discuss. For example, our sources have revealed that Sun was working on derivatives of the multi-core Niagara chip that would function as co-processors for Rock. Such chips could be used to speed up certain high performance computing workloads and give Sun an edge over rivals. During the recent round of layoffs, Sun cut back on one of the main co-processor projects dubbed Jupiter.

A Sun labs staffer declined to say what the company now plans to use on the co-processor front, although he did confirm that Sun expects some manner of co-processor in the DARPA system.

The US hopes that Sun, IBM and Cray will deliver designs that stay well ahead of Japanese and eventually Chinese supercomputing efforts for years to come.

Embedded Slowlaris

Solaris on PowerPC? Yes, we've been there before, but Sun is banging away at the idea again.

A number of people went at putting Solaris on Power back in the mid-1990s with not too much to show from their efforts.

"This idea has a different feel now though," said Tom Riddle, a coder in Sun Labs.

Riddle is angling Solaris on PowerPC at the embedded market with embedded standing for anything that isn't a general purpose computer or server. Solaris 10, in particular, is well suited for the embedded crowd due to its stability and a wealth of features beyond those in Linux that can be tapped by developers.

The basic port work still needs to be completed. From there, Riddle hopes to create a smaller, lightweight kernel, some new power management tools and real-time function improvements that will make Solaris better suited for the embedded space.

There's a lot of momentum behind PowerPC right now and a decent amount behind Solaris 10, so we'll see how this plays out. Although, it's hard to see how Solaris offers a huge edge over Linux in the embedded market.

Labbing around

The latest visit to Sun Labs comes after Sun announced 5,000 redundancies and vowed to cut back on non-essential spending. Close to 200 workers make up Sun Labs, and they consume close to 2 per cent of Sun's annual $2bn R&D spend.

The likes of Sun, IBM and HP (ignoring Itanium and Mark Hurd) deserve tremendous credit for sticking with their ambitious R&D programs during a business era that would prefer they give up on innovation. The US has an incredible track record of turning R&D - funded by the government and corporations - into major new businesses. This tradition complements the use of venture capital.

It's hard to imagine how the US can continue to compete against the low wages and abundance of engineers in Asia without relying on its R&D and venture capital strengths. Some element of the American character thrives in the R&D-backed entrepreneurial arena.

Seeing fellows such as Bob Sproull and Ivan Sutherland at Sun Labs provides some hope that US companies can push forward on the R&D front with great success. ®