From: "Luck, Tony" <tony.luck@intel.com>
To: Peter Zijlstra <peterz@infradead.org>,
        Vikas Shivappa <vikas.shivappa@linux.intel.com>
CC: "Shivappa, Vikas" <vikas.shivappa@intel.com>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        "x86@kernel.org" <x86@kernel.org>, "hpa@zytor.com" <hpa@zytor.com>,
        "tglx@linutronix.de" <tglx@linutronix.de>,
        "mingo@kernel.org" <mingo@kernel.org>,
        "Shankar, Ravi V" <ravi.v.shankar@intel.com>,
        "Yu, Fenghua" <fenghua.yu@intel.com>,
        "Anvin, H Peter" <h.peter.anvin@intel.com>
Subject: RE: [PATCH 4/6] x86/mbm: Memory bandwidth monitoring event
 management
Thread-Topic: [PATCH 4/6] x86/mbm: Memory bandwidth monitoring event
 management
Thread-Index: AQHRdBTRYrill/rqkU+ff/3h6/+MD59PKRKA//99WGA=
Date: Mon, 7 Mar 2016 23:27:26 +0000
Message-ID: <3908561D78D1C84285E8C5FCA982C28F3A000C1D@ORSMSX114.amr.corp.intel.com>
References: <1456876108-28770-1-git-send-email-vikas.shivappa@linux.intel.com>
 <1456876108-28770-5-git-send-email-vikas.shivappa@linux.intel.com>
 <20160307230329.GT6344@twins.programming.kicks-ass.net>
In-Reply-To: <20160307230329.GT6344@twins.programming.kicks-ass.net>
Accept-Language: en-US
Content-Language: en-US
Content-Type: text/plain; charset="us-ascii"
Content-Transfer-Encoding: 8BIT
MIME-Version: 1.0
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1208
Lines: 28

>> +		bytes = mbm_current->interval_bytes * MSEC_PER_SEC;
>> +		do_div(bytes, diff_time);
>> +		mbm_current->bandwidth = bytes;
>> +		mbm_current->interval_bytes = 0;
>> +		mbm_current->interval_start = cur_time;
>> +	}
>>> +
>> +	return mbm_current;
>> +}
>
> How does the above time tracking deal with the event not actually having
> been scheduled the whole time?

That's been the topic of a few philosophical debates ... what exactly are
we trying to say when we report that a process has a "memory bandwidth"
of, say, 1523 MBytes/s?  We need to know both the amount of data moved
and to pick an interval to measure and divide by. Does it make a difference
whether the process voluntarily gave up the cpu for some part of the interval
(by blocking on I/O)? Or did the scheduler time-slice it out to run other jobs?

The above code gives the average bandwidth across the last interval
(with a minimum interval size of 100ms to avoid craziness with rounding
errors on exceptionally tiny intervals). Some folks apparently want to get
a "rate" directly from perf.  I think many folks will find the "bytes" counters
more helpful (where they control the sample interval with '-I" flag to perf
utility).

-Tony