In this installment of the Best of the Beowulf Mailing List we review some postings about noise issues (always a problem) and clock resolution for code profiling. You can consult the archives for the actual conversations.

Beowulf: Quiet and Powerful

If you have stood next to a rack of 1U dual processor computers you know how loud they can be. On December 12, 2003, Joshua Baker-LePain asked about systems that provided the best bang for the noise (performance per dB anyone?). He currently has six 1U nodes with dual 2.4 GHz Xeon processors in them. Each node has eleven small fans, resulting in a setup that was very loud for the people around them (and there were people in the same room as the machines). He is looking for systems that are much quieter but without a penalty in speed. Nicholas Henke was the first to respond that his desktop Dell 650N box was very quiet and it was a dual processor machine.

Beowulf: Quiet and Powerful

If you have stood next to a rack of 1U dual processor computers you know how loud they can be. On December 12, 2003, Joshua Baker-LePain asked about systems that provided the best bang for the noise (performance per dB anyone?). He currently has six 1U nodes with dual 2.4 GHz Xeon processors in them. Each node has eleven small fans, resulting in a setup that was very loud for the people around them (and there were people in the same room as the machines). He is looking for systems that are much quieter but without a penalty in speed. Nicholas Henke was the first to respond that his desktop Dell 650N box was very quiet and it was a dual processor machine.

Bill Broadley also chimed in that he has a Dell 360N that he felt was quieter than the Dell 650N. He also said that he's found the single P4 processor in the 360N to be faster than the 650N even running two processes, because the memory subsystem in the 360N was substantially better then the 650N. He also mentioned that he's never heard a quiet 1U or 2U system. He's found that even Apple's XServe systems to be loud. He finally recommended that if you are building a system yourself, to use rubber grommets in the case and to use large slow spinning 120 mm fans.

Joe Landman took the same idea of Bill's and put it in general terms. He recommended using larger size fans since they move more air for a given RPM (revolutions per minute). This decision usually dictates a larger case which Joe also mentioned allows for sound absorption to be installed fairly easily. He offered that a 2U or 3U case of good quality might be able to be cooled effectively while remaining quiet and allowing for sound absorption to be installed without adversely affecting airflow.

Greg Lindahl also offered an explanation for the noise. The typical 1U case uses very small fans that are inefficient. So a large number of fans have to be used to achieve the required cooling and airflow. Dean Johnson added to Greg's comment by stating that the smaller fans in the 1U cases run at a higher RPM to improve the airflow which raises the pitch of the noise. In Greg's words, this "... pegs the annoy-o-meter." He gave an example where he had an SGI 1100 (a 1U, dual PIII node) and an SGI Origin 200 (a regular case size system running MIPS processor(s) ). He said that both probably had about the same level of loudness, but he turned off the 1100 when he wasn't using it because of the pitch of the noise.

Daniel Pfenniger offered a link to a company with a desktop box that s designed for low noise. Daniel said that he had one in his office that could barely be heard. He thought that putting even 12 in a room would not disturb a conversation. (26 dBa)

Andrew Latham also mentioned that koolance has water cooling kits for standard rack mount cases that could potentially reduce the noise level. However, Joel Jaeggli chimed in that in fact the case was a 4U case. He said that in a 4U of space, he could get eight Opteron 242's (four dual setups) with panaflo crossflow blowers in each which he felt was quieter than the small 40mm, 8,000 RPM fans in typical 1U cases.

In the drive for speed, system vendors increased the packaging density of systems from the old 4U cases, to 2U, to 1U, and finally to blade servers. At the same time, the power requirements, and thusly the heat dissipation requirements, for the systems had gone up. Trying to effectively cool them and quiet them is a monumental task. This brief exchange on the beowulf mailing list points out there are times when quiet systems are really needed.