<div class="gmail_quote">On 20 January 2011 16:45, pl bossart <<a href="mailto:bossart.nospam@gmail.com" target="_blank">bossart.nospam@gmail.com</a>> wrote: <blockquote class="gmail_quote" style="margin: 0pt 0pt 0pt 0.8ex; border-left: 1px solid rgb(204, 204, 204); padding-left: 1ex;"> <div>> Along with increasing the buffer size, the tsched wakeup watermark will need > to be increased, although my initial prototype code did not do that. I > waited until the watermark increased to the point where there were no more > underruns before using powertop. On my PC, it seems like the process for > decreasing the watermark is very aggressive, maybe to the point of causing > more underruns than needed. > > I plan on doing more testing, but on an ARM based platform, obviously, I am > looking for a buffer growth tuning for ARM that will reduce CPU wakes. For > dynamic growth, I am thinking that with the check_left_to_play (alsa-sink.c) > when woken by a timeout, it would be a good time to try to increase buffer > size and hopefully allowing ARM to go to a lower power state on long plays. > The buffer would be set back to normal default when rewinding, etc. </div>You can't really change the size of the buffer, it's defined when you define the hw_params. You can only change the timer value. It think the 'buffer growth' you want is already handled by the concept of latency. We've also experimented with larger buffers. It's not clear to me that you really benefit from 8+ sec buffers. You have other wake-ups in the system that will reduce the benefits of long sleeps. Also as you said the timer watermark needs to take the tlength into account, a fixed initial value isn't very good. And last there is still a wake every 1.5s due to the auto-timing update. It would need to be changed to depend on the latency value. It's been on my TODO list but I still don't understand how the smoother parameters need to be changed. -Pierre </blockquote><div> </div></div>Thanks for your reply Pierre. I am still learning and so I spent some time this weekend experimenting. I tried again to show larger buffer size would help, but there was really very little difference, just as you suggested. I then tried to change the buffer size during a playback, when we woke on_timeout in alsa_sink.c and snd_pcm_hw_params failed. In fact, due to the failure, the module was actually unloaded. I will do more experimentation on watermark reduction and auto-timing update with relationship to latency and learn that code better. I was wondering if you could elaborate on what you had on your TODO list for this area with regards to smoother parameters. I am following the dynamic sample rate discussion as that also may help CPU on ARM. Any other comments and suggestions of areas to look at would also be appreciated. Kurt Taylor (irc krtaylor)