6.8rc1 Tunables

[WizADSL]

is there any guidance available for MD tunables on version 6.8rc1?  Does the size or number of discs or system memory affect what the best values would be?  In the presence of very high I/O I still have some performance issues and I would like to see if any of the parameters can be adjusted to mitigate that.

[John_M]

I asked about two of them. Here's Tom's reply:

https://forums.unraid.net/bug-reports/prereleases/unraid-os-version-680-rc1-available-r631/page/3/?tab=comments#comment-5723

 

[WizADSL]

Thank you for that.  I'm hoping someone can chime in on the remaining parameters.

Edited October 15, 2019 by WizADSL
Spelling

[JorgeB]

You can turn the help on, it gives an idea, but basically and as far as I understand they were introduced to fix the writes starving reads problem with v6.7, first one controls the max queue depth for each separate md device, from 1 to 100%, so with default value a single device can use 80% of the entire strip pool, if there's I/O on two devices at the same time each one can use 40%, and so on, should always be less than 100.

 

md_sync_limit limits the amount of the stripe pool parity sync/check is permitted to consume when there are active I/O streams on other md devices (if there is no other I/O parity sync/check is permitted to consume up to md_queue_limit), also expressed as a percentage from 0..100, suggestion from LT is to keep it low.

 

 

 

 

Edited October 15, 2019 by johnnie.black

[WizADSL]

On 6.8rc1 I am still seeing writes starve reads.  I' wondering if these parameters are an issue for me since other have reported that 6.8rc1 has solved the problem for them.

[JorgeB]

Just now, WizADSL said:

On 6.8rc1 I am still seeing writes starve reads.

First one reporting still having the issue, so possibly unrelated to original one, but make sure you're using the default settings, they work for everybody else, if still having issues describe how to reproduce.