These servers were connected to a managed switched Gigabit Ethernet network via DLink switches, then to our controller system which was an HP Notebook running Windows Vista SP1. We monitored power consumption using a Raritan Dominion PX PDU with 8 outlets which were directly read from an external meter. In turn, we calibrated and monitored the Raritan unit with a WattsUP?stet Pro Power Meter and its software through a USB port on the aforementioned HP notebook.
We tested four servers, an IBM x3350, a Dell 1950, an HP DL160 and an HP DL360.
The server hardware components of the IBM x3350 were one Quad-Core Intel Xeon 2.66GHz CPU; 4GB of DIMMS memory; dual integrated Gigabit Ethernet ports; and, two 250GB 7.2K SATA hard disk drives using a LSI Logic SAS/SATA Controller.
The server components for the Dell 1950 were two Quad-Core Intel Xeon 2.50GHz CPUs; 16GB of DIMM memory; 73GB hard disk drives using a PERC integrated SAS/SATA daughtercard controller; and two embedded Broadcom NetXtreme II 5708 Gigabit Ethernet network interface cards.
The server components for the HP DL160 G5 were Quad-Core Xeon 3.00GHz CPU, 4GB PC2-5300 fully buffered DIMMs with Advanced ECC 2 support, an embedded NC105i PCIe Gigabit LSI Logic Smart Array E200/128MB Controller attached to two160GB 1.5G SATA 7.2K hard disks.
The server components for the HP DL360 G5 with dual Quad-Core Xeon 3.00GHz CPU, 12GB PC2-5300 fully buffered DIMMs, two embedded NC105i PCIe Gigabit LSI Logic Smart Array E200/128MB Controllers attached to four 72GB 1.5G SATA 7.2K hard disks (configured as RAID 5).
These servers were connected to a managed switched Gigabit Ethernet network via DLink switches, then to our controller system which was an HP Notebook running Windows Vista SP1. We monitored power consumption using a Raritan Dominion PX PDU with eight outlets that were directly read from an external meter. In turn, we calibrated and monitored the Raritan unit with a WattsUP? Pro Power Meter and its software through a USB port on the aforementioned HP notebook. We didn’t start taking power consumption until after the initial power draw cycle was over. The initial draw of power each server made was often in excess of 50% increase over the power consumed by a booted operating system.
Quiescent State Testing
We first installed each operating system on each server, set each server for balanced/performance settings, and we measured four hours of quiescent activity. We then reset the power savings to the maximum settings we could find, and retested four hours activity.
Active State Testing
We reset all servers and operating systems to the balanced/performance configuration and settings and then installed mail software on each server to measure power draw on an active server. Sendmail/procmail with 1,000 LDAP-linked users was installed on both SUSE and Red Hat in separate tests conducted in sequence. Alternately, for the Windows 2008 test we installed Microsoft Exchange Server with the same 1,000 users imported into Active Directory for this test. During each mail test, we bombarded the mail servers with e-mail targeted to the users continuously; the e-mails were generated from several concurrently running scripts to generate the continuous activity.
We then restarted the servers with maximum power savings settings/configurations as used in the quiescent tests in place, and tested again.
Measurement Protocol
Each server under test was assigned to a specific outlet on the Raritan PDU. The PDU was plugged into and powered through the watt meter, which had a USB connection to the laptop that ran the WattsUP?Pro realtime logging software. For each test run, the meter logging software was set to take the total (global) power draw (including the PDU's draw), voltage, current, etc. samples every 15 seconds.
In parallel a shell script was run to retrieve the information available (power, current, voltage, etc.) from each individual outlet through SNMP. The script was configured to retrieve the information from each outlet in round robin fashion every 15 seconds so that an individual outlet would have its information sampled every 45 seconds. Additionally, the PDU's information (power, current, voltage) is collected along with each outlet's information to correlate with the external watt meter's readings. PDU overhead is included in the measurement.
< Return to main story: Linux captures the “green" flag, beats Windows 2008 power saving measures >