Hitting the (memory) wall

One of the issues that comes with increasing the number of cores in a processor is memory access.  Back when AMD initially came out with their dual-core processors they were WAY better than the Intel models, in large part because their chipset allowed for faster access to the memory.  That gave them a pretty huge in-road in the processor market for servers and a solid couple month head start over Intel in the multi-core processor battles.

Now chip manufacturers are starting to worry about memory bottlenecks in upcoming generations of multi-core processors.  Recent testing at Sandia National Labs suggests that going past an 8-core processor actually causes problems with memory access.  A 16-core processor actually performed at about the same rate as a dual-core processor.  Oops. 

Apparently Intel knows about the problem and they’re working on it, but it is still somewhat worrisome that the future of virtualization and other fun improvements in system performance and functionality pretty much depends on this working.  Processor speeds are slower now than they used to be and future growth is going to be in the number of cores, not processor speed.  If the performance issues can’t be solved then we’re in big trouble.

The Hypervisor doesn’t matter

Most presentations from company reps focus on just how great their product is and how much it is what all the others are not.  When I hear a company man saying something else it is certainly refreshing, and it happened again recently in a conversation with a Senior Engineer on the Windows Server team.  His responsibility pretty much covers the virtualization space, and this is the second time I’ve heard him talk about Hyper-V (notes from the first session here, here and here).  And just like last time this session was a very insightful look at the Microsoft product as well as the direction the industry is heading overall. 

The most poignant comment that was made throughout the session was that the industry has had far too much focus on the hypervisor and not enough on the actual value of virtualization.  The vendors have all pretty much agreed that the hypervisor itself is a commodity – there just aren’t that many differences between them and any differences that there are will be wiped out in the coming months as the vendors continue to improve their products.  The key is to look past server consolidation towards the areas where virtualization can truly start improving the performance of an environment, not just cut down on the number of physical servers in your environment. 

So how will that happen?  Part of it is more efficient and functional management tools.  The ability to manage workloads rather than just servers (i.e. shift a workload to a different server to account for a hung process within a single server) is a huge component of this next step in management tools improvements.  The other big aspect of improved management tools is efforts that are being made to improve the server deployment process.  Today’s process is pretty much to deploy a sysprep’d server image and then spend a couple hours adding various components and configuration settings and then finally loading the server application(s) on the system.  Microsoft is putting efforts in to streamlining the deployment process.  So you can indicate that a virtual server will be an Exchange 2007 Hub Transport Server and it can know to install the .net framework and other various components as part of the provisioning process.  This idea of “composing” a server rather than installing apps is going to be a major piece of Microsoft’s push in their management suite and application installation process going forward.

Of course, I’m not all that surprised to hear Microsoft calling the hypervisor a component, particularly since theirs isn’t one of the best out there right now in terms of overall functionality (e.g. no live system migration/Vmotion, though supposedly that is functional but didn’t make the cut for the RTM release).  Still, the move towards actually making better use of the the virtual environment beyond just reducing physical server count is a rather notable goal.

Some other minor notes from the session:

• Don’t expect any further cooperation between AMD and Intel to allow guest OS sessions to migrate between servers running the different processor families anytime soon.  Not likely to happen.
• Multi-core VMs are almost always a waste because so few applications are written to support multi-threaded operation.  Nothing new here, but it is worth repeating.
• The next version of Windows Server probably won’t scale up all that much more in terms of processor core support.  Today’s 64-core support seems to actually be sufficient for 98%+ of the server hardware or implementation needs in the world.  The next version might up the number to 128 or 256 cores, but there really isn’t all that much of a need for it as best as anyone can see.
• Server consolidation doesn’t do much for server application availability.  Remember that next time you’re trying to sell the benefits of virtualization to someone and they ask what the benefits are.

More notes from the session next week.  In the mean time, start getting used to the idea that the hypervisors are all pretty much commodities and that the management aspects are the most important thing to look at when you’re planning a new environment.  And you’re ultimately probably going to want to mix and match to get the best benefits from all the vendors.

XP SP3 notes beginning to trickle out

Now that XP SP3 has been out for a couple weeks, some details are starting to trickle out regarding some of the bumps in the road associated with the deployment.

The first issue is related to AMD-powered HP desktops that are running the HP image of Windows XP. Like many of us, HP uses a single image across all their hardware, both AMD and Intel PCs. The image loads a driver for Intel power management, which is fine right up until SP3 is installed on an AMD-powered PC and the system reboots. At that point the system tries to use Intel codes on the AMD motherboard and things do not end well. Disabling the driver in advance of the SP3 install will solve the issue, but not knowing that in advance is a recipe for a very unhappy upgrade.

Another issue is related to a change in the Internet Explorer security zones. Of course, the effect has absolutely nothing to do with Internet browsing and everything to do with copying files from a network location to a local PC. The issue is similar to the effect of running menu links from a shared network location, where a security prompt would come up if the network location isn't in your trusted sites list. There are some steps for a workaround, and they can be managed via GPOs, but it is annoying.

There are some other fun improvements that XP SP3 offers, including for Microsoft's implementation of Network Access Protection, so it is definitely worth installing. But, like everything else, make sure to test it well before deploying to users.

Microsoft's vision of virtualization -- Part 2

As promised in yesterday's post, here is part 2 of my summary from the Microsoft Virtualization thing I attended this week, focusing on desktop virtualization technologies and where Microsoft sees them going.

Microsoft has made three significant purchases in the past couple years that focus on the desktop virtualization arena. There was the Softricity purchase for application virtualization, and a new release of that is coming soon (v4.5) as part of the Microsoft Desktop Optimization Pack (MDOP) that should almost be Enterprise-ready for streaming applications to an existing OS install. More significant than that, however were the purchases of Calista and Kidaro, two companies that focus on performance and functionality for Virtual Desktop Infrastructure (VDI) implementations.

In the server virtualization space densities of 20 or 50 guests to 1 host are pretty good. But if you're trying to get a VDI implementation for a couple thousand workstations a 50:1 ratio still doesn't meet the needs of most organizations. There is a lot of effort going in to improving the hypervisor performance to meet these needs, but it is not quite there yet. As for the Calista and Kidaro purchases, I now have a much better understanding of what they are all about.

• Calista was focused almost completely on improving graphics performance in the RDP stream. They had made huge strides in reverse engineering the protocol and then leveraging the GPU on the local machine, rather than the CPU, to make graphics performance in a terminal session much, much, much better. With the recent claim by someone from nVidia that the GPU is more important than the CPU in terms of future improvements in system performance, the focus on using the GPU is actually completely reasonable. From Microsoft's perspective the Calista purchase is all about improving the RDP protocol's graphics handling though any means they can leverage. Combining this with the synthetic drivers functionality that I mentioned yesterday where the full feature set of the hardware can be leveraged through the hypervisor, and things are looking up in the graphics arena for terminal services shops. Still, there is some acknowledgement that neither RDP nor ICA are able to provide the full experience of device connectivity, UI and performance that a true VDI solution would require. That makes it reasonably likely that the Calista technology will get rolled into the new super-RDP, whenever that comes out, to make VDI more feasible.
• The Kidaro purchase is a cross between Virtual PC and Published Applications in a Citrix/Terminal Server environment. Kidaro includes both a client and a server component and allows for an application to run completely in a Virtual PC environment on a local PC with the appearance of just a regular local application. Plus, the server component allows for the VHD associated with the app to be managed centrally and distributed or revoked from the console, precluding the need to individually manage the VHDs on all the PCs. So if you have some legacy application or custom code that requires Office 95 to function but you also want to deploy Vista you can build a centrally managed Windows XP SP3 VHD and install Office 95 on it, distribute the VHD via Kidaro to the appropriate people and also provide a seamless UI experience via Kidaro. The effect will be similar to how VMware's Fusion product works with Macs, allowing a seamless window for a Windows app to run inside OS X. For folks who have worked in a terminal services environment and struggled with the seamless versus window-in-a-window UI this is particularly appealing. The Kidaro product will be rolled into the MDOP once it is fully integrated into the Microsoft suite.

Microsoft is also continuing to work with Citrix/Xen on some connection brokering functionality. The Citrix platform has been well ahead of Microsoft on the connection brokering (published apps, load balancing, etc.) and web interface pieces for several years now, so Microsoft is continuing to leverage that, in combination with their Virtual Machine Manager product (more on this tomorrow) and Hyper-V platform. The ability to have individually assigned VM sessions for some users who have special needs or the ability to provide an generic "shared" VM for users with similar needs allows for scalability and functionality similar to a terminal services environment, but with isolation between the users. This all depends on a new approach to "profile virtualization" to maintain user settings. In reality this is just folder redirection, but using virtualization in the name of something makes it cooler, right?

A real VDI solution that scales and provides a rich user experience is still a couple years out, but in certain niche markets, like call center apps, it will happen much sooner.

Dell pulls back on AMD

Barely 18 months after bringing AMD into the fold, Dell is pulling back on the chip vendor, removing almost every desktop model using the processor from their website. Dell claims that the demographics don't match up, with AMD being mostly used by consumers outside the United States, and Dell mostly selling to domestic companies. They are keeping the AMD processors on their server line, which is where AMD has a more compelling play in the business world anyways, so it isn't that big a deal, but the press is eating it up.