Message-ID: <adb4042b-4fd9-4a64-5904-b00c4ae8d473@linux.vnet.ibm.com>
Date:   Tue, 31 Jan 2023 21:20:53 +0530
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.6.1
Subject: Re: [RFC PATCH] hrtimer: interleave timers for improved single thread
 performance at low utilization
Content-Language: en-US
To:     Arjan van de Ven <arjan@linux.intel.com>, tglx@linutronix.de
Cc:     peterz@infradead.org, mingo@kernel.org,
        Srikar Dronamraju <srikar@linux.vnet.ibm.com>,
        svaidy@linux.ibm.com, linux-kernel@vger.kernel.org,
        bigeasy@linutronix.de
References: <5ae3cb09-8c9a-11e8-75a7-cc774d9bc283@linux.vnet.ibm.com>
 <c6c106a2-b180-d3fd-5904-44f9b4949ddc@linux.intel.com>
From:   shrikanth hegde <sshegde@linux.vnet.ibm.com>
In-Reply-To: <c6c106a2-b180-d3fd-5904-44f9b4949ddc@linux.intel.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Precedence: bulk


On 1/31/23 8:25 PM, Arjan van de Ven wrote:
> 
>>   kernel/time/hrtimer.c | 11 +++++++++++
>>   1 file changed, 11 insertions(+)
>>
>> diff --git a/kernel/time/hrtimer.c b/kernel/time/hrtimer.c
>> index 3ae661ab6260..d160f49f0cce 100644
>> --- a/kernel/time/hrtimer.c
>> +++ b/kernel/time/hrtimer.c
>> @@ -1055,6 +1055,17 @@ u64 hrtimer_forward(struct hrtimer *timer,
>> ktime_t now, ktime_t interval)
>>
>>           orun = ktime_divns(delta, incr);
>>           hrtimer_add_expires_ns(timer, incr * orun);
>> +        /*
>> +         * Avoid timer round-off, so that all cfs bandwidth timers
>> +         * don't start at the same time
> 
> so while I applaud the final objective, I am sort of wondering if
> hrtimer.c is the right place in the kernel to fix a CFS/cgroup issue...
> wouldn't it be better to solve such issues at the place we want this to
> happen, rather than for all timers in the whole system?
> 
Agree. It was an initial approach to the problem. This can be fixed in scheduler 
code itself as suggested by Ingo. Will send out that patch soon. 
> (also while for performance it might be better to spread out a bit, for
> power consumption it's obviously the other way around)
> 
This would be performance vs power comparison. Would get both the 
numbers next time for fairer comparison.