lkml.org 
[lkml]   [2018]   [Apr]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 1/6] x86/intel_rdt/mba_sc: Add documentation for MBA software controller
On Thu, 29 Mar 2018, Vikas Shivappa wrote:
> +Memory bandwidth(b/w) in MegaBytes
> +----------------------------------
> +
> +Memory bandwidth is a core specific mechanism which means that when the
> +Memory b/w percentage is specified in the schemata per package it
> +actually is applied on a per core basis via IA32_MBA_THRTL_MSR
> +interface. This may lead to confusion in scenarios below:
> +
> +1. User may not see increase in actual b/w when percentage values are
> + increased:
> +
> +This can occur when aggregate L2 external b/w is more than L3 external
> +b/w. Consider an SKL SKU with 24 cores on a package and where L2
> +external b/w is 10GBps (hence aggregate L2 external b/w is 240GBps) and
> +L3 external b/w is 100GBps. Now a workload with '20 threads, having 50%
> +b/w, each consuming 5GBps' consumes the max L3 b/w of 100GBps although
> +the percentage value specified is only 50% << 100%. Hence increasing
> +the b/w percentage will not yeild any more b/w. This is because
> +although the L2 external b/w still has capacity, the L3 external b/w
> +is fully used. Also note that this would be dependent on number of
> +cores the benchmark is run on.
> +
> +2. Same b/w percentage may mean different actual b/w depending on # of
> + threads:
> +
> +For the same SKU in #1, a 'single thread, with 10% b/w' and '4 thread,
> +with 10% b/w' can consume upto 10GBps and 40GBps although they have same
> +percentage b/w of 10%. This is simply because as threads start using
> +more cores in an rdtgroup, the actual b/w may increase or vary although
> +user specified b/w percentage is same.
> +
> +In order to mitigate this and make the interface more user friendly, we
> +can let the user specify the max bandwidth per rdtgroup in bytes(or mega
> +bytes). The kernel underneath would use a software feedback mechanism or
> +a "Software Controller" which reads the actual b/w using MBM counters
> +and adjust the memowy bandwidth percentages to ensure the "actual b/w
> +< user b/w".
> +
> +The legacy behaviour is default and user can switch to the "MBA software
> +controller" mode using a mount option 'mba_MB'.

You said above:

> This may lead to confusion in scenarios below:

Reading the blurb after that creates even more confusion than being
helpful.

First of all this information should not be under the section 'Memory
bandwidth in MB/s'.

Also please write bandwidth. The weird acronym b/w (band per width???) is
really not increasing legibility.

What you really want is a general section about memory bandwidth allocation
where you explain the technical background in purely technical terms w/o
fairy tale mode. Technical descriptions have to be factual and not
'could/may/would'.

If I decode the above correctly then the current percentage based
implementation was buggy from the very beginning in several ways.

Now the obvious question which is in no way answered by the cover letter is
why the current percentage based implementation cannot be fixed and we need
some feedback driven magic to achieve that. I assume you spent some brain
cycles on that question, so it would be really helpful if you shared that.

If I understand it correctly then the problem is that the throttling
mechanism is per core and affects the L2 external bandwidth.

Is this really per core? What about hyper threads. Both threads have that
MSR. How is that working?

The L2 external bandwidth is higher than the L3 external bandwidth.

Is there any information available from CPUID or whatever source which
allows us to retrieve the bandwidth ratio or the absolute maximum
bandwidth per level?

What's also missing from your explanation is how that feedback loop behaves
under different workloads.

Is this assuming that the involved threads/cpus actually try to utilize
the bandwidth completely?

What happens if the threads/cpus are only using a small set because they
are idle or their computations are mostly cache local and do not need
external bandwidth? Looking at the implementation I don't see how that is
taken into account.

Thanks,

tglx






\
 
 \ /
  Last update: 2018-04-03 11:47    [W:0.182 / U:1.848 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site