Posted by: Mark Fontecchio
blade server, HP BladeSystem
LAS VEGAS — Today I attended a session at the HP Technology Forum on powering blade servers, and heard some interesting data points on how much system configuration matters to power consumption.
Tony Harvey, HP product manager for blade server power and cooling, gave one particularly interesting example. In it, he compared two configurations:
- Configuration 1: 16 blade servers using Intel quad-core 2GHz E5335 processors, 8 GB RAM (8 x 1GB), an additional dual-port Ethernet, a Cisco Ethernet switch and a Brocade FibreChannel switch.
- Configuration 2: 16 blade servers using Intel quad-core 2GHz L5335 processors, 8 GB RAM (2 x 4GB), an additional dual-port Ethernet, an HP Ethernet switch and an HP Fibre switch.
On the surface, these look like practically identical configurations. And yet the first configuration uses 30% more power at 100% utilization, and about 15% more power when idle. Why? Because of the subtle differences:
- Both use a quad-core Intel processor running at 2GHz, but the L5335 uses less power than the E5335.
- Using two 4GB DIMMs instead of eight 1GB DIMMs is more efficient
- Harvey said using the HP switches is more efficient than the Cisco and Brocade versions.
It doesn’t end there. Harvey continued on, saying the application you run also affects your blade servers’ power consumption. HP tested a configuration of 16 blades running various applications, and found that power consumption varied from about 4 kilowatts for a rendering application, to more than 7 kilowatts for Linpack, one of the most power-hungry application there is.
Earlier in the presentation, Harvey mentioned that a full rack of blade servers — that’s four chasses — could consume as much as 35 kilowatts. I asked him afterward what that configuration would include:
- 32 BL 2×220 server nodes per chassis/enclosure
- 32 GB RAM per server
- Infiniband in every blade
- 2 Infiniband switches
- Run Linpack and smoke that baby! (Note: Harvey never said “smoke that baby!” That was me.)
Is that a realistic production configuration? Harvey said no. The only time he’s ever seen that kind of configuration is when a company like HP is trying to score well on the Linpack benchmark, or when they’re doing a burn-in to stress-test the systems.