Hi Peter, >Peter ZhaoKai wrote:
>Dear Tesla Team, > >Nice to meet you. This is Kai Zhao (Peter), I am wondering what to contribute >and > what I could do. I do test automation and QA works in my job. Nice to meet you, too, :) > >This project seems an edge-scientific case, it saves the power, how much >electricity > power we could save for our next generation ? :-) > I'm not sure Eric posted this onto osnews, but recently we reached a big milestone, Power Aware Dispatcher and Deep Idle driver has been integrated in the Solaris kernel, working together with the P-state driver, I'm pretty sure the power consumption of this planet will be reduced by what we did, ;p > >My newbie opinions, >In the "Power Observability" that probes be used to send running staticis into >to PM for > analyzising. besides cpu-change-speed and idle-state-transition, these parts > could also > be probed on monitoring in the DTrace Power Provider: > Virtual memory page swap in and out frequency, Disk I/O amount (Perhaps OLTP > will > consumes much CPU workload to do db transaction on > elect/update/delete/insert), > RAM data refresh frequency or even CPU cache refresh frequency, workload from > the > network adaptor. Yeah, these events definitely eat power. Currently we focus on cpupm, PowerTOP is a software utility designed to measure and explain a computer's electrical power consumption. The observable information is used to pinpoint software, device driver and kernel option that result in excessive power use and could cover those events you mentioned, but just stand on the CPU's point of view. Of course there are much opportunities we can do for power observability. > >Some other foolish opinions: The workload between GPU and CPU, nowdays GPU are >powerful, >CPU calculation power could be saved ? I'm not the graphics expert, but I believe the solaris graphics driver is driving GPU to do as much and most efficient as possible to manipulate and display computer graphics. >And if there are 3 CPU, one for processing long-time tasks, one for short but >frequent tasks, >the last one could have a rest. That's exactly what Power Aware Dispatcher is doing, but more complicated than your thoughts. If you have interest, here is a good description and the source code you can read. http://www.opensolaris.org/os/project/tesla/Work/CPUPM/ > >BTW, how DTrace probes monitor several multi-core CPU ? DTrace probes don't do any active job, they are passive actually. For example, the probe "idle-state-transition" exists at the per-cpu code path, so when a cpu is going to idle, the probe will be hitted, and there is a magic at this point to dynamically replace several instructions to generate a trap, that leads the code path into the DTrace handler, so that we can export any info we need. > >Thanks, >Peter Thanks, -Aubrey
