Re: [PATCH 5/6] selftests/resctrl: Do not compare performance counters and resctrl at low bandwidth

[Ilpo Järvinen]

On Thu, 5 Sep 2024, Reinette Chatre wrote:
> On 9/5/24 4:45 AM, Ilpo Järvinen wrote:
> > On Wed, 4 Sep 2024, Reinette Chatre wrote:
> > > On 9/4/24 4:43 AM, Ilpo Järvinen wrote:
> > > > On Fri, 30 Aug 2024, Reinette Chatre wrote:
> > > > > On 8/30/24 4:42 AM, Ilpo Järvinen wrote:
> > > > > > On Thu, 29 Aug 2024, Reinette Chatre wrote:
> > > > > > 
> > > > > > > The MBA test incrementally throttles memory bandwidth, each time
> > > > > > > followed by a comparison between the memory bandwidth observed
> > > > > > > by the performance counters and resctrl respectively.
> > > > > > > 
> > > > > > > While a comparison between performance counters and resctrl is
> > > > > > > generally appropriate, they do not have an identical view of
> > > > > > > memory bandwidth. For example RAS features or memory performance
> > > > > > > features that generate memory traffic may drive accesses that are
> > > > > > > counted differently by performance counters and MBM respectively,
> > > > > > > for instance generating "overhead" traffic which is not counted
> > > > > > > against any specific RMID. As a ratio, this different view of
> > > > > > > memory
> > > > > > > bandwidth becomes more apparent at low memory bandwidths.
> > > > > > 
> > > > > > Interesting.
> > > > > > 
> > > > > > I did some time back prototype with a change to MBM test such that
> > > > > > instead
> > > > > > of using once=false I changed fill_buf to be able to run N passes
> > > > > > through
> > > > > > the buffer which allowed me to know how many reads were performed by
> > > > > > the
> > > > > > benchmark. This yielded numerical difference between all those 3
> > > > > > values
> > > > > > (# of reads, MBM, perf) which also varied from arch to another so it
> > > > > > didn't end up making an usable test.
> > > > > > 
> > > > > > I guess I now have an explanation for at least a part of the
> > > > > > differences.
> > > > > > 
> > > > > > > It is not practical to enable/disable the various features that
> > > > > > > may generate memory bandwidth to give performance counters and
> > > > > > > resctrl an identical view. Instead, do not compare performance
> > > > > > > counters and resctrl view of memory bandwidth when the memory
> > > > > > > bandwidth is low.
> > > > > > > 
> > > > > > > Bandwidth throttling behaves differently across platforms
> > > > > > > so it is not appropriate to drop measurement data simply based
> > > > > > > on the throttling level. Instead, use a threshold of 750MiB
> > > > > > > that has been observed to support adequate comparison between
> > > > > > > performance counters and resctrl.
> > > > > > > 
> > > > > > > Signed-off-by: Reinette Chatre <reinette.chatre@xxxxxxxxx>
> > > > > > > ---
> > > > > > >     tools/testing/selftests/resctrl/mba_test.c | 7 +++++++
> > > > > > >     tools/testing/selftests/resctrl/resctrl.h  | 6 ++++++
> > > > > > >     2 files changed, 13 insertions(+)
> > > > > > > 
> > > > > > > diff --git a/tools/testing/selftests/resctrl/mba_test.c
> > > > > > > b/tools/testing/selftests/resctrl/mba_test.c
> > > > > > > index cad473b81a64..204b9ac4b108 100644
> > > > > > > --- a/tools/testing/selftests/resctrl/mba_test.c
> > > > > > > +++ b/tools/testing/selftests/resctrl/mba_test.c
> > > > > > > @@ -96,6 +96,13 @@ static bool show_mba_info(unsigned long
> > > > > > > *bw_imc,
> > > > > > > unsigned long *bw_resc)
> > > > > > >       		avg_bw_imc = sum_bw_imc / (NUM_OF_RUNS - 1);
> > > > > > >     		avg_bw_resc = sum_bw_resc / (NUM_OF_RUNS - 1);
> > > > > > > +		if (avg_bw_imc < THROTTLE_THRESHOLD || avg_bw_resc <
> > > > > > > THROTTLE_THRESHOLD) {
> > > > > > > +			ksft_print_msg("Bandwidth below threshold (%d
> > > > > > > MiB).
> > > > > > > Dropping results from MBA schemata %u.\n",
> > > > > > > +					THROTTLE_THRESHOLD,
> > > > > > > +					ALLOCATION_MAX -
> > > > > > > ALLOCATION_STEP *
> > > > > > > allocation);
> > > > > > 
> > > > > > The second one too should be %d.
> > > > > > 
> > > > > 
> > > > > hmmm ... I intended to have it be consistent with the ksft_print_msg()
> > > > > that
> > > > > follows. Perhaps allocation can be made unsigned instead?
> > > > 
> > > > If you go that way, then it would be good to make the related defines
> > > > and
> > > > allocation in mba_setup() unsigned too, although the latter is a bit
> > > > scary
> > > 
> > > Sure, will look into that.
> > > 
> > > > because it does allocation -= ALLOCATION_STEP which could underflow if
> > > > the
> > > > defines are ever changed.
> > > > 
> > > 
> > > Is this not already covered in the following check:
> > > 	if (allocation < ALLOCATION_MIN || allocation > ALLOCATION_MAX)
> > > 		return END_OF_TESTS;
> > > 
> > > We are talking about hardcoded constants though.
> > 
> > Borderline yes. It is "covered" by the allocation > ALLOCATION_MAX but
> > it's also very non-intuitive to let the value underflow and then check for
> > that with the > operator.
> 
> My understanding is that this is the traditional way of checking overflow
> (or more accurately wraparound). There are several examples of this pattern
> in the kernel and it is also the pattern that I understand Linus referred
> to as "traditional" in [1]. Even the compiler's intrinsic overflow checkers
> do checking in this way (perform the subtraction and then check if it
> overflowed) [2].

Fair enough. I've never come across that kind of claim before.

> > You're correct that they're constants so one would need to tweak the
> > source to end up into this condition in the first place.
> > 
> > Perhaps I'm being overly pendantic here but I in general don't like
> > leaving trappy and non-obvious logic like that lying around because one
> > day one of such will come back biting.
> 
> It is not clear to me what is "trappy" about this. The current checks will
> catch the wraparound if somebody changes ALLOCATION_STEP in your scenario, no?
>
> > So, if a variable is unsigned and we ever do subtraction (or adding
> > negative numbers to it), I'd prefer additional check to prevent ever
> > underflowing it unexpectedly. Or leave them signed.
> 
> To support checking at the time of subtraction we either need to change the
> flow of that function since it cannot exit with failure if that subtraction
> fails because of overflow/wraparound, or we need to introduce more state that
> will be an additional check that the existing
> "if (allocation < ALLOCATION_MIN || allocation > ALLOCATION_MAX)"
> would have caught anyway.
> 
> In either case, to do this checking at the time of subtraction the ideal way
> would be to use check_sub_overflow() ... but it again does exactly what
> you find to be non-intuitive since the implementation in
> tools/include/linux/overflow.h uses the gcc intrinsics that does subtraction
> first and then checks if overflow occurred.

It's trappy because by glance, that check looks unnecessary since 
allocation starts from max and goes downwards. Also worth to note,
the check is not immediately after the decrement but done on the next 
iteration.

The risk here is that somebody removes allocation > ALLOCATION_MAX check.

Something called check_sub_overflow() is not subject to a similar risk 
regardless of what operations occur inside it.

> It is not clear to me what problem you are aiming to solve here. If the
> major concern is that the current logic is not obvious, perhaps it can
> be clarified with a comment as below:
> 
>  	if (runs_per_allocation++ != 0)
>  		return 0;
>  +	/*
> +	 * Do not attempt allocation outside valid range. Safeguard
> +	 * against any potential wraparound caused by subtraction.
> +	 */
>  	if (allocation < ALLOCATION_MIN || allocation > ALLOCATION_MAX)
>  		return END_OF_TESTS;

That would probably help but then it seems Linus is against such attempts
and considers this hole in the cheese (i.e., representing something that 
is clearly a negative number with a positive number) "traditional".

-- 
 i.