Is "Because I can" a requirement?
Sun has had a bit of a history of making product decisions based on what we think customers should have as opposed to what they want. I call that the "broccoli versus candy" syndrome. Sure, broccoli is better for you but it is a lot easier to sell candy. Fortunately that has changed, and Sun's top management and product teams are much more in tune with what needs to be done to win the ground war. That might sound obvious but there is a delicate balance between timing the market and stamping out current-for-a-day products.
One product illustrates this particularly well: the V20z. The V20z was a unique product when it came out. With two dual core AMD Opteron CPUs and HyperTransport in a single rack unit server, it offered about the highest CPU power per rack unit available. The value was all in the packaging. It gave better performance, used less power, less cooling and less real estate than anything with similar compute power.
The problem was that we still lost a lot of V20z sales because customers didn't seem to have a requirement for such a package.
Case in point. I had a meeting with a customer when the V20z and V40z came out. That customer was used to just buying packs of Dell servers off the web, five or ten at a time to have them on hand. In this case we were comparing the Dell 2850 to the V20z. The V20z was the clear winner in so many respects: performance, power consumption, cooling, rack space. The Xeon CPUs in the 2850 were no match for the AMD architecture. The fly in the ointment for the customer was the number of drives. The Dell supported six drives, albeit in double the rack space, where the V20z supported two drives. The customer stated that he always bought the servers with all drive slots occupied. "Well," I thought to myself, " Let's see what the requirement is."
"So you get six drives in all the machines? What do you do with them?" I ask. The admin responds,
"I mirror the first two 73 gig drives as a boot volume, then put three 73 gig drives in a parity group and the last 73 gig drive is a hot spare for the other two volumes."
"Ok, wow, so how much space do you need for applications and data on these servers?" I said.
"We only use about ten to twenty gigs." Now I am getting confused. I ask,
"Umm, so is it a matter of spindles or capacity?"
"Neither," says the admin. "I can get six drives in the Dell so that's what I buy."
That is when I learned that "Because I can" seems to be a legitimate business requirement. No matter how inexpensive the Dell may have been, the customer was literally throwing money away. By any count, including price, the V20z would have done the job perfectly well. The two drives could be mirrored and still would have provided way more performance and capacity than necessary. To boot, it would have done it with more CPU power, a third of the electricity and half the rack space. The customer, however, didn't have any requirements surrounding power, cooling or rack space. We were selling broccoli and this customer wanted candy.
Another bit of irony with the V20z and V40z was that they were among the most powerful "PC" servers on the market at a time when the industry wasn't quite sure what to do with extremely powerful PC servers. The operating systems usually associated with small PC servers don't scale well past two or four cores, unlike Solaris which happily scales on machines with 128 cores.
As I mentioned earlier, things are very different today, a few years after the introduction of the V20z and V40z. The current line of X64 rack and blade servers are as impressive as ever, and I'm happy to explain how they trump the competition.
Oh, and about selling candy instead of broccoli? Ask us for an NDA Roadmap presentation, we love to do them. The upcoming servers will have plenty of drives.
Things have not improved much...
Last week I was looking at ordering a SPARC Enterprise M4000 server. A key reason for considering this machine is for workloads that need high single-threaded performance but not a lot of threads. I can domain it down to two machines, right?
If you take a closer look at Sun SPARC Enterprise M4000/M5000/M8000/M9000 Servers Administration Guide you will see that domain 0 gets all (there are only two) disks, all of the Ethernet ports, and 3 of the 5 PCI/PCIe slots. This makes it so that if I need to connect domain 1 to a SAN, my only option is to put a dual or quad ethernet card in one slot and a dual-port HBA in another. It forces me to boot from the SAN (or NFS, I guess).
What would a customer that would be likely to by a significant number of these be looking for?
How much would have that changed the cost of a $100k machine? My guess is less than 2%, even with the drive bays populated. Sure, PCI bridges may have turned out a bit different and space for another SAS and Ethernet chip would likely need to be found. Big deal.
This leaves me with the option of using half-populated M4000's or using multiple domains on half-populated M5000's. The first option means I am using 3U per CPU, and the second option takes me to 2.5U per CPU. I guess I won't have to worry about the cooling problems that high density servers bring.
Before suggesting alternative technologies like zones, rest assured I put zones to use where they are appropriate. Zones are not up to the task in this and many other cases because I need total resource isolation and the ability to run different OS or patch releases on the domains.
FWIW, the reason I didn't buy V20z's was because I needed redundant power supplies.
Posted by Mike Gerdts on July 23, 2007 at 08:42 AM EDT #