Fwd: [PATCH] Hitshield : Something new eviction process for MGLRU

---------- Forwarded message ---------
보낸사람: Minwoo Jo [off-list ref]
Date: 2024년 10월 9일 (수) 오전 1:04
Subject: Re: [PATCH] Hitshield : Something new eviction process for MGLRU
To: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: <akpm@linux-foundation.org>, <rostedt@goodmis.org>,
[off-list ref], [off-list ref], [off-list ref],
[off-list ref], [off-list ref], [off-list ref],
[off-list ref], [off-list ref]


2024년 10월 8일 (화) 오후 10:21, Mathieu Desnoyers
[off-list ref]님이 작성:

On 2024-10-08 15:02, Mathieu Desnoyers wrote:

quoted

On 2024-10-08 10:44, Minwoo Jo wrote:

quoted

Signed-off-by: Minwoo Jo <redacted>

[ Any reason why no mailing lists are CC'd ? ]

[...]

quoted

Please consider CCing LKML and the linux-mm mailing lists on your reply.

The Documentation/process/submitting-patches.rst document describes who
should be CC'd when submitting patches.

The script scripts/get_maintainer.pl will help you identify which
email recipients are relevant for your patch.

Thanks,

Mathieu

--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com

[ Any reason why no mailing lists are CC'd ? ]

I think I did some mistake when I mail, so I add CC to
linux-mm@kvack.org (open list:MEMORY MANAGEMENT)
linux-kernel@vger.kernel.org (open list)
linux-trace-kernel@vger.kernel.org (open list:TRACING)

Have you considered alternative approaches, and why is this approach
better ?

I used an alternative approach of creating another wrapper or value
approach, but this was not good because it increased memory usage per
folio, resulting in more swaps.

Do you have benchmarks of workloads that regress with this feature
enabled ? How are they affected ?

I tried another two benchmarks, 7zip benchmark, YCSB.

The 7zip benchmark shows a 3.1% reduction in pswpin, 2.8% reduction in pswpout,
and a 2.9% reduction in runtime under 1G constraints.
Under 2G constraints, pswpin decreased by 22.4%, pswpout decreased by 21.2%,
and Runtime decreased by 14.8%.
On 3G constraints, pswpin decreased by 3.3%, pswpout decreased by 1.3%,
and runtime decreased by 1.9%.

We also measured the Yahoo! Cloud Service Benchmark.
YCSB showed a 69.9% decrease in pswpin, 67.1% decrease in pswpout,
and a 5.1% decrease in Runtime at the 500M constraint.
At 750M constraints, pswpin decreased by 83.8%, pswpout decreased by 81.0%,
and runtime decreased by 2.4%.
On 1G constraints, pswpin decreased by 79.5%, pswpout decreased by 76.7%,
and runtime decreased by 3.8%.

Documentation should be there first, or at least something explaining
the rationale and design.

I think my previous commit mail might be helpful.
https://lore.kernel.org/linux-mm/ZpvY_aZV4VtMmXP-@casper.infradead.org/T/ (local)

Thank you for your response.