A Novella: Replace the Dell Racks with new hardware
Subtitle: A win-win solution proposed by Physical Plant and ITS

Once upon a time, back in 2013, two Dell racks full of compute nodes, sat noisily chewing away energy on the 5th floor of Science Tower. They drew in nicely cooled air from the floor spewing it out the back of the racks at 105-110 degrees (F). They were giving the three Liebert cooling towers a run for their BTUs. So much so that if one failed the Dell racks needed to be powered down to avoid the data center reaching temperatures beyond 95 degrees (F). The Dell racks were in a foul mood ever since that last event, not too long ago. And so, day after day, they consumed lots of BTUs, and with the ample supply of Watts coming from their L6-30 roots, converted it all into heat. Tons of heat, making life lousy for the Liebert family. Oh, and they performed some computational work too, but if even they did not, the energy consumption remained the same. That's a fact. They were 6 years old and determined to make it to 12. So the story goes.

The Dell racks contain 30 compute nodes, two UPS units, two disks arrays and two switches. We have measured 19 nodes power consumption (pulling one of the dual power units out) with a Kill-A-Watt meter for over 775+ total hours. The mean power consumption rate is 418.4 watts. That totals to 109,956 KwH/year in power consumption ((watts/1000 Kw per hour) * 24 hours * 365 days * 30 servers). This is a low water mark, it only takes into account the compute nodes but that will be the majority of heat producers. We also measured one rack's consumption at the utility panel and Peter's calculation yields 126,000 KwH/year which can be considered a high water mark.

Next we need to convert to a dollar value. A residential electric bill contains a KwH cost as well as generation, distribution, transmission and other cost factors. Typically the total KwH cost is 50% of the total bill. Utility power in the data center comes from our cogen plant as well as CP&L. The model for cogen attempts to balance both; 15% from CP&L and 85% cogen generated. If the cogen plant is down, like for maintenance, this jumps to 100% CP&L. One also needs to factor in natural gas costs, heat reclamation costs, etc. If we set the overall cost at 12.5 cents “per KwH” we're assuming 6.25 cents per cogen generated KwH consumed, which seems reasonable. This value is set by Peter.

Based on 12.5 cents the Dell compute nodes consume $13,744.50 per year in power. Best guess is cooling costs are at least that (another possible low water mark). So the total cost for both power and cooling consumption is $27,489 per year.

Next step was to collect vendor quotes for a target budget of $82K, 3 years of Dell energy consumption, an arbitrary length of time. That's so we can downscale from there because the new racks of course still consume energy. Four quotes were obtained and they show a similar pattern. Here is the comparison given key features.

Old hardware: 109,956 KwH/year for power
30 nodes, 2.66 ghz, 4 mb L-cache (for cpu), 240 cores (job slots),
80 gb local drive, 340 gb total ram, 12,555 watts (power no cooling), 670 gigaflops (actual measure)

New hardware v1: 47,304 KwH/year for power or 43% of old hardware
14 nodes, 2.60 ghz, 20 mb L-cache (for cpu), 224 cores (job slots),
1TB local drive, 1,792 gb total ram, 5,400 watts (power no cooling), 4,659 gigaflops (theoretical)

New hardware v2 (half of v1): 23,652 KwH/year for cooling or 22% of Old hardware
7 nodes, 2.60 ghz, 20 mb L-cache (for cpu), 112 cores (job slots),
1TB local drive, 1,792 gb total ram, 2,700 watts (power no cooling), 2,329 gigaflops (theoretical)

If we reduced the node count to 7 (the minimum configuration to meet the job slot count of the Dell hardware), the total energy consumption (power plus cooling) would be 5,400 watts. The total cost of running the new hardware (v2) would be $5,913 per year. That would imply savings of $21,576 per year. And that's using the low water mark. The job slot count would be 112 but with hyperthreading technology that can be doubled. We'd still want the 1,792 memory footprint (8 gb/core) and the gigaflops (2,329) still far exceeds Dell's performance.

In two years, the new hardware would have saved $43,152 on energy costs based on the low water mark (Dell's costs would equal $55K). We still need to adjust some minor issues:

• There are enough Infiniband ports available to put all new hardware nodes on such a switch (add cards and cables cost for each node)
• The internal disks on each node need to be of a high speed (10K or better) and of a certain size (300 GB or larger) mimicking the Dell disk arrays (adds costs)
• we maybe able to add two more nodes by switching to a more exapansive lower wattage CPU (and remain within budget as well as below the 50% energy consumption threshold as compared with Dell's consumption.
  • accomplished by switching from 8 core 2650v2 (130 watt) 2.6 ghz CPU to 10 core 2660v2 (95 watt) 2.2 ghz CPU

But it is all very doable within a budget of $45-$50K. And it can be the solution for:

• replace Dell's racks functions and match or exceed its performance
• seriously reduce energy consumption benefiting Physical Plant's bottom line
• allow ITS to treat the third Liebert cooling tower as backup/standby generating more energy savings
• being way green

The Libert family rejoices. The Dell family moves out. The end.

The table below contains data for a cluster whose nodes are all on the Infiniband switch (and also ethernet switch for provision and data). They also contain a 15K 300 GB SAS drives each for access to local fast disk (Gaussian users). It still deployes the 8-core CPUs, thus 16 pysical cores per node, 32 hyperthreaded cores per node and in both cases 256 GB of memory.

The Dell racks were bought in 2006. They contain 30 compute nodes, two UPS units, two disks arrays and two switches. Measurements of 2/3rds of the compute nodes with a Kill-A-Watt meter yields an average consumption of 418.4 watts (if the nodes are computing or not). That totals to 109,956 KwH/year for power, a low water mark. Measurements at the utility panel for one of the racks yields a consumption of 126,000 KwH/year. Cooling requirements (not measured) are assumed to be equal to that.

Using the low water mark, and a cost per KwH (inclusive of cogen generation costs, maintenance, CP&L power imports, heat reclamation costs, etc), the total cost for both power and cooling consumption is estimated at $27,489 per year for the Dell racks.

New hardware could replace the Dell's functionality and reduce power/cooling needs in the data center while yielding significant savings. An 8-node cluster with each node comprised of dual 8 core CPUs, each with 256 GB of memory and a 15K 300 GB hard disk, all nodes connected to a high throughput/low latency switch, would match or exceed key parameters such as gigaflops of computational power and number of job slots provided (with hyperthreading enabled).

Such a cluster would consume 77% less energy generating $21,094 in saving per year (after accounting for energy needs of that 8-node cluster). This implies that in 2.9 years the cost of acquiring that 8-node cluster ($61,714) will be recouped. Based on the low water mark.

Back