by Nik Simpson | July 9, 2010 | 1 Comment
As Intel and AMD have announced new products, the energy efficiency of their processors in terms of performance/watt has been a key part of their overall marketing message. But the energy efficiency of the processor is only a small part of the overall efficiency of an infrastructure designed to support applications, and often the processor is tiny part of the picture that includes storage, memory, networking equipment, and the overall efficiency of the power distribution in the data center.
So this week’s announcement of the first TPC database benchmark results that include energy consumption as a metric were of great interest. The results come from HP and cover their recently announced 4-socket servers; DL580 G7 (Intel Xeon 7500 family) and D585 G7 (AMD Opteron 6000 family). From my perspective, the most interesting results are the TPC-E results that allow us to compare the DL 580 and DL 585 directly on the same benchmark:
The configurations used were similar in construction, but the DL580 used considerably more hardware:
The extra hardware means that the DL580 G7 uses substantially more power, but delivers 43% transactions than DL585 G7, but on a transaction/watt basis the DL580 G7 is ~13% more efficient.
What’s more striking is how little the server contributes to the overall power consumption for a database application, as we can see from the following table (taken from the Executive Summary for the DL585 G7 benchmark submission):
Not surprisingly, the real energy hog here is the storage component which accounts for almost 80% of the energy consumption. Even halving the energy consumption of the server would have marginal impact on the overall picture. So if you want to make database systems more energy efficient, focus on the storage, and that means looking at technologies such as solid state disks.
So, congratulations to the TPC and HP for getting these results out, I’m really looking forward to how other vendors will respond, and how quickly we see some innovation in the storage subsystems. One last point that I would like to see addressed is that reporting energy consumption is optional, and that needs to change, not just for the TPC benchmark suite but for all server and storage benchmarks.
Posted by: Nik Simpson
Category: Benchmarks energy efficiency Servers Storage Tags:
by Nik Simpson | June 29, 2010 | Comments Off
Last week I took a brief look at two companies who believe they have the solution to energy efficient clouds (see here). Both approaches certainly sound like they should deliver more energy efficient clouds than can be constructed using conventional commodity servers, but that may not be enough, for several reasons:
- Niche solution: Both Tilerra and SeaMicro’s approaches only address a part of the cloud compute infrastructure market, specifically platform-as-service clouds. They are not suitable for hardware-infrastructure-as-a-service clouds which must provide x86 compatibility and the flexibility to run different size workloads. SeaMicro’s approach gets a tick for the first condition, but can’t support virtual machine workloads that require more compute power than a single Atom CPU can deliver. Tilerra fails on both counts, in that it lacks x86 compatibility and uses processor cores that offer modest individual performance.
- Vendor stability: For a company like Google, this isn’t a major problem, if they decided that such an architecture is the way to go, then they can pick up either company for what would amount to small change for them. But for less established players, particularly VC funded startups this would be a deal breaker. I’ve worked with VCs in the past, and they are unlikely to back a startup that has to rely on the success of another startup’s product.
- Established vendors: The established vendors are increasingly focused on the needs of cloud service providers for two reasons. First, they represent a significant market opportunity since they buy servers in the tens of thousands. Second, if cloud displaces conventional IT as many predict, then cloud providers may become the primary market. So I expect that we’ll see a lot innovation from the established vendors in this area, and their manufacturing economies of scale will be tough to beat.
So, Kudos to Tilerra and SeaMicro for innovative products, but it’s going to take a lot more than elegant engineering to build a business here, and I’m not sure that either company fully comprehends that.
Posted by: Nik Simpson
Category: Cloud Servers Tags: cloud hardware architectures, energy efficiency
by Nik Simpson | June 23, 2010 | 1 Comment
I’ve touched on the possibility of using non-traditional hardware architectures such as ARM-based servers in previous posts on the Burton Group blog (see here, here) but some recent announcements from AMD, SeaMicro, and Tilera are starting to make a clearer case for cloud providers.
First up is this week’s announcement from AMD of its Opteron 4000 line of processors for dual socket servers. AMD’s major claim for these processors is that they are better suited to large scale cloud deployments than Intel’s Nehalem family of processors because they are more power efficient. A quick look at the relevant specs does back this claim up, with the Xeon 5600 bottoming out at 40W/processor vs.. 32W for the lowest power version of the Opteron 4000, a 20% improvement for AMD.
However, both SeaMicro and Tilera feel that increasing power efficiency of Opteron and Nehalem class processors is simply putting lipstick on a power hog. Both companies take the position that conventional x86 architectures are inherently inefficient with regard to power consumption for scale-out cloud infrastructures because they are overkill. The problem is compounded by all the supporting logic such as network, storage, graphics, and other components that have to be on the motherboard to create a useful system. Each component adds a few watts to the power consumption, and when you are deploying thousands of servers those watts add up to some pretty hefty power bills.
Though SeaMicro and Tilera are singing from the same hymn book, their designs are very different. SeaMicro has chosen to develop a system around Intel’s Atom processor to maintain x86 compatibility, but that’s as far as the resemblance to conventional x86-based server goes. The system consists of multiple mini-servers in a 10U chassis. Each mini-server consists of an Atom processors, Intel’s standard Atom PCIe chipset, 2GB of non-ECC memory, and SeaMicro’s ASIC that contains all the supporting logic in a single package including high-speed interconnects to network and storage resources.Eight mini-servers are integrated onto a single plug-in card with 64 cards in each 10U chassis. Using this approach they are able to pack 2048 mini-servers and 4 TB of memory into standard rack with a power footprint in the 8 kW range.
Tilera goes a step further, creating a complete system on a chip that packs 64 processing cores (not x86 compatible), and supporting logic such as networking into a single chip, Using this system, their design partner Quanta is building a skinless 2U design that holds up to eight Tilera chips, supporting 512 cores and 256 GB of memory (4GB DIMMs) in 2U. That corresponds to 10,752 cores and 5.24 TB of memory in a single rack with a power footprint in the 10 kW range.
The challenge for both SeaMicro and Tilera is to gain to mindshare in a very competitive market where system vendors such as Dell are creating servers specifically optimized for large cloud deployments and some of the cloud providers buy servers in such quantities that they can go direct to pacific rim manufacturers and get servers built to their precise specification. I’ll touch on some of the challenges and opportunities for SeaMicro and Tilera in a future blog post.
Posted by: Nik Simpson
Category: Cloud Servers Tags: