Since Dell launched the PowerEdge server line 10 years ago, we have introduced servers in waves or “generations”, where each generation is defined by a common set of system technologies—common networking chips, RAID subsystems, etc. The approach makes it easier for customers to adopt those technologies and has been well received, but concentrating change into narrow windows does emphasize those changes. A couple of weeks ago, Dell introduced its 9G PowerEdge servers, an event that for me highlighted a series of changes—both for Dell and the industry in general. On a personal level, it marked the end of four years of managing the server development team. Since the introductions, I’ve moved on to lead platform engineering for all Dell products, along with my “two-in-a-box” partner Stuart Caffey.
From an industry perspective, we are seeing a real shift in historical trends. For years, system performance was one of the most important considerations for vendors and our customers, with power and thermal impacts to datacenters farther down the list of priorities. Over the past 5 years we’ve seen huge performance increases, but power has increased as well. Citing a Dell example—other companies are similar—back in 2000, our mainstream 2U server utilized a 330-watt power supply. Just four years later, power requirements for a similar server more than doubled to 700 watts. Rising energy costs and the increasingly complex issue of re-working datacenters have made power and cooling a priority for customers. I’ve spoken with some customers who are contemplating building new datacenters that panicked when they looked at historical power curves and projected them into the future. Some system companies have taken advantage of that panic and offered some “power solutions” of highly dubious value like introducing low-voltage (48V) DC power distribution into data centers—not a good idea for a lot of reasons. The good news for the folks planning datacenters is that the industry has listened and power trends have hit an inflection point: the mainstream servers of 2006 will dissipate less power than their predecessors while delivering far more performance.
Back to the 2U example, the latest Dell 2U, the PowerEdge 2950 with the Intel 5100 series processor, delivers more than double the overall performance but requires 25% less power at maximum load. The trend changed because the whole industry, from the system companies to the processor vendors, realized the problem a few years ago and shifted our thinking around to “power in the datacenter is a fixed, precious resource” from a “if we build it they will come, no matter what the power/cooling costs.”
I don’t think we’ve seen the high water mark in power and say that future servers will always dissipate lower power. Quad-core processors are coming in the next year and they will draw more power than today’s CPUs. You can see the foreshadowing of quad core in the power supply ratings on some systems just introduced: the ratings are higher than needed for the current round of CPUs, memory, etc. What we’re seeing is the end of power doubling—if power goes up it will likely be minimal going forward. Dell, and likely others, will focus on reducing power and improving efficiency. One focus will be on offsetting increases in one area with decreases in another. A subtle example in this generation of Dell servers is the use of TCP/IP Offload Engines (TOE) in our embedded Ethernet network adaptors. Dell made TOE standard on our mainstream servers in part because it can improve network performance, but mainly because it greatly reduces CPU utilization and hence power in many real-world applications.
We’d like to hear your thoughts. Stay tuned for more details from Dell on this topic.
Update: Dell one2one reader David G. had a comment about versarails system. In this vlog, Greg Henderson, an engineer from our server rack team, shows off the functionality.