• Virtualization has overhead, but probably a lot less than the 43-50% sometimes claimed.
• Just to be safe, don’t virtualize apps that are already I/O-bound or otherwise running flat-out. (So there’s no contradiction to my support for dedicated security, networking, and data warehouse appliances.)
• Big enterprises have lots of production servers that are old, unreliable, and/or idle most of the time. Virtualize those.
• If a server’s use is particularly spiky, it may be a great candidate for virtualization.
• Most development servers can and should be virtualized.

Makes sense to me.

Subsequent to submitting the column, I developed a simpler taxonomy of computing appliance types, namely:

Type 0: Custom hardware including proprietary ASICs or FPGAs.

Type 1: Custom assembly from off-the-shelf parts. In this model, the only unusual (but still off-the-shelf) parts are usually in the area of network acceleration (or occasionally encryption). Also, the box may be balanced differently than standard systems, in terms of compute power and/or reliability.

Type 2 (Virtual): We don’t need no stinkin’ custom hardware. In this model, the only “appliancy” features are in the areas of easy deployment, custom operating systems, and/or preconfigured hardware.

Here’s what I predict for each of them.

Type 0 appliances used to be favored by a range of networking and security vendors, such as Netscreen. However, there’s a trend away from them. Netezza is the only Type 0 appliance vendor I’ve actually talked with recently.

Type 1 appliances are pretty much the norm, as per the column and summary post linked above. However, that depends in no small part on how networking-intensive the appliance’s software is. Interestingly, DATAllegro seems to be moving from Type 1 to Type 2.

Type 2 is where most appliance vendors ideally would want to be. Examples can already be found in data warehousing, antispam, and firewall. Also, a variety of platform vendors are interested in supporting virtual appliances, including VMware, Microsoft, maybe IBM, and Crossbeam. Crossbeam’s strategy may be the most interesting of all.

Software appliances and SW configurations integrated with virtual middleware

You can see the whole list here. Here’s more about Ferguson and his role.

Half or more of the computing appliance vendors I’ve looked into follow very similar hardware strategies: They use mainly standard parts; they include uncommon but off-the-shelf networking (and sometimes encryption) accelerators; and they of course optimize the mix of those parts and general hardware architecture as well. (EDIT: I actually gave names to three strategies — even if they were just “Type 0″, “Type 1″, and “Type 2″ — in this overview of data warehouse appliance vendors. And in another post I considered arguments about whether one would want a data warehouse appliance at all.) Examples I’ve posted about recently include – and I quote the forthcoming column – “DATallegro and Teradata (data warehousing), Cast Iron Systems (data integration), Barracuda Networks (security/antispam), Blue Coat Systems (networking), and Juniper (security and networking).” (ANOTHER EDIT: But I think DATAllegro’s strategy has changed.)

By way of contrast, there’s also a group whose stance is more along “hardware/schmardware” lines. Sendio and Proofpoint (in most cases) don’t really do anything special at all in their boxes; what’s more, Proofpoint actually has significant software-only deployments over VMware’s virtualization layer. Kognitio and Greenplum think their software-only data warehouse offerings are appliance-equivalents too; indeed, Greenplum’s software is sold mainly bundled with Sun hardware (to the extent it’s sold at all), and Kognitio is hinting at an appliance-like offering for competitive reasons as well. Check Point Software plays both sides of the field; it offers its own kind of “virtual appliance,” but also gets many of its sales through appliance vendors. Its most interesting such partner, if not its biggest, is Crossbeam Systems, which in my opinion may very well represent the future of appliance technology.

• Demo (of course).
• Tweak, test, deploy – between patches and new anti-spam rulesets, Proofpoint users seem to have a rapid change/test/deploy cycle. Virtualization makes it possible to do that without having multiple copies of an appliance.
• Disaster recovery – this seems to be a big one.
• “Surges” – depending on what the bad guys are doing, one’s need for anti-spam servers can go up and down in a hurry. Virtualization makes it easy to respond.

It probably is not coincidental that Proofpoint makes less use of custom hardware that many other appliance vendors. In most cases, Proofpoint just buys servers from a vendor such as Dell and fiddles with their packaging. For some applications it does add enhanced networking capabilities, as other appliance vendors do; while I neglected to ask explicitly, I got the impression these weren’t the ones commonly deployed on VMware.

Proofpoint said that very few of its customers bought VMware in connection with their Proofpoint deployments; except in a couple of cases, they already had it and were using it for other things.

• VMware quotes a figure of 20,000 enterprise customers, if you count everybody who is at least testing the software and so on; i.e., it’s a somewhat inflated figure. Still, the “real” number is surely big.
• They claim ¼ of those have a “VMware first” policy, to deploy new apps on a virtual rather than dedicated machine. That’s impressive until you realize enterprise try to roll their own apps as rarely as possible these days.
• They suggest VMware is extremely helpful at times you’d like to have two copies of the same platform – e.g., for development, or when you have to take the system down for brief maintenance. It’s hard to argue with that claim.
• We didn’t have the time to talk about my performance concerns.

As for how this all plays with appliances and SaaS – that’s largely a future, but potentially a very interesting one. Here’s what I mean.

In July, 2005, VMware starting pushing the concept of “virtual appliance” with 6 ISVs, mainly large software vendors who wanted to make their software easy to deploy for evaluation cycles. By November, 2006 they’d rolled out a “marketplace” with 380 ISVs, although many of those seem to be single individuals offering software for free. Microsoft, Cisco and so on have rival announcements of partnerships and so on, but I haven’t ascertained whether these get past the Barney level (“I love you, you love me” without real substance).

Basically, a virtual appliance is a bundled software stack (e.g., DBMS plus OS), with more or less the same ease-of-deployment advantages boasted by regular appliances (assuming, of course, that you’ve deployed virtualization in the first place). However, virtual appliances lack the other big advantage of regular appliances – specialized hardware. Right now, they do not support the networking cards, encryption chips, and so forth that many appliance vendors use. What’s more, as I noted previously, I don’t see how they could be as well tuned in their hardware/software combinations as real appliances are. VMware has APIs and a “community source” program to encourage one-off extensions to support special hardware, but mainly this is an area where future innovation will be needed. As to when that innovation will come, and whether it will come primarily from the hardware or the virtualization side — well, that remains to be seen. But if I wanted to grow market share or start a new venture on the hardware side today, I’d be looking to generic appliance support (with a strong blade orientation, of course) as the way to make my mark.

And here’s something that very few people seem to have thought about so far – virtual SaaS. As with appliances, SaaS has as one of it major selling points ease of deployment. What’s more, another of its big virtues – rapid update cycle – requires super-easy deployment as well. Traditionally, this is done on a hosted basis. But there’s actually no reason the same benefits couldn’t be provided using virtual appliance technology. So if issues of data privacy or control or whatever are barriers to the adoption of SaaS applications – well, they won’t always have to be. Conversely, if traditional packaged software vendors want to co-opt SaaS’s benefits, they should be perfectly able to, performance-overhead-cost issues perhaps aside.

EDIT: The specific VMware gentleman I spoke with was

• Srinivas Krishnamurti
• Director, Product Management and Market Development

]]> http://www.monashreport.com/2007/01/03/virtual-appliances-saas/feed/ 4 http://www.monashreport.com/2006/07/15/appliances-are-not-dead-yet/ http://www.monashreport.com/2006/07/15/appliances-are-not-dead-yet/#comments Sat, 15 Jul 2006 06:30:09 +0000 Curt Monash http://www.monashreport.com/2006/07/15/appliances-are-not-dead-yet/ Nick Carr and Jonathon Schwartz are predicting the death (or at least decline) of special-purpose computing appliances. Their reasons, so far as I can tell, are pretty much threefold:

1. Vendors have economies of scale making general-purpose computers.
2. Users have economies of scale running homogenous, general-purpose computers.
3. Virtualization will work.

But when one thinks a little bit about what’s really driving the use of appliances, those arguments fall apart.

Throughout the history of commercial computing, special purpose data center devices have succeeded in two areas — communications, and high-volume storage — for three main reasons:

1. These uses involve moving lots of data, at high speeds. Unnecessary overhead hurts.
2. These uses are generally simple and repetitive. Much of the overhead of flexible, general-purpose computing is unnecessary.
3. For internet devices only, most especially ones that do security.) Single-purpose devices are much easier to lock down.

Since the mid-1990s, communication/networking has been exploding as a fraction of overall computing — and therefore, the same has been true of appliances.

If you don’t think this trend will continue, then probably either:

• You think Layer 7 inspection will reverse it, because a device that can do Layer 7 inspection needs to be able to run general-purpose apps.
• You think virtualization will reverse it, because virtualization lets hardware be optimized for multiple kinds of uses at once.

But I think both of those opinions would be wildly overstated.

A good illustration of these points may be found in an exchange I had with Stuart Frost, CEO of DATallegro. DATallegro is a data warehouse appliance maker heavily optimized for streaming data on and off disk, moving it between mulitiple processors, and making the best use of onboard cache, memory bus, etc. Stuart thinks it’s likely that he could get much of the same benefit from shipping on a known configuration of name-brand equipment. But I can’t imagine any way that DATallegro’s software would work well on any kind of virtualized multi-purpose grid kind of setup. Blech. The vast majority of their engineering would be simply wasted.

Nor will Cisco routers be replaced by software running on Sun grids any time soon.