Re: [PATCH bpf-next v2 0/5] execmem_alloc for BPF programs

From: Song Liu <song@kernel.org>
To: "Edgecombe, Rick P" <rick.p.edgecombe@intel.com>
Cc: "mcgrof@kernel.org" <mcgrof@kernel.org>,
	"peterz@infradead.org" <peterz@infradead.org>,
	"bpf@vger.kernel.org" <bpf@vger.kernel.org>,
	"rppt@kernel.org" <rppt@kernel.org>,
	"dave@stgolabs.net" <dave@stgolabs.net>,
	"Torvalds, Linus" <torvalds@linux-foundation.org>,
	"linux-mm@kvack.org" <linux-mm@kvack.org>,
	"hch@lst.de" <hch@lst.de>, "x86@kernel.org" <x86@kernel.org>,
	"akpm@linux-foundation.org" <akpm@linux-foundation.org>,
	"Lu, Aaron" <aaron.lu@intel.com>
Subject: Re: [PATCH bpf-next v2 0/5] execmem_alloc for BPF programs
Date: Tue, 8 Nov 2022 10:20:07 -0800	[thread overview]
Message-ID: <CAPhsuW55Rb+=7SawWUOM9OEcdjHfSAkT2CkknBgWj+5HPhAE0A@mail.gmail.com> (raw)
In-Reply-To: <dc47953aa9296d1955e41f02d4ddef06036d855c.camel@intel.com>

On Mon, Nov 7, 2022 at 4:13 PM Edgecombe, Rick P
<rick.p.edgecombe@intel.com> wrote:
>
> On Mon, 2022-11-07 at 15:39 -0800, Luis Chamberlain wrote:
> > On Mon, Nov 07, 2022 at 03:13:59PM -0800, Song Liu wrote:
> > > The benchmark used here is identical on our web service, which runs
> > > on
> > > many many servers, so it represents the workload that we care a
> > > lot.
> > > Unfortunately, it is not possible to run it out of our data
> > > centers.
> >
> > I am not asking for that, I am asking for you to pick any similar
> > benchark which can run in paralellel which may yield similar results.
> >
> > > We can build some artificial workloads and probably get much higher
> > > performance improvements. But these workload may not represent real
> > > world use cases.
> >
> > You can very likely use some existing benchmark.
> >
> > The direct map fragmentation stuff doesn't require much effort, as
> > was demonstrated by Aaron, you can easily do that or more by
> > running all selftests or just the test_bpf. This I buy.
> >
> > I'm not buying the iTLB gains as I can't even reproduce them myself
> > for
> > eBPF JIT, but I tested against iTLB when using eBPF JIT, perhaps you
> > mean iTLB gains for other memory intensive applications running in
> > tandem?
>
> Song, didn't you find that there wasn't (or in the noise) iTLB gains?
> What is this about visible performance drop from iTLB misses?
>
> IIRC there was a test done where progpack mapped things at 4k, but in
> 2MB chunks, so it would re-use pages like the 2MB mapped version. And
> it didn't see much improvement over the 2MB mapped version. Did I
> remember that wrong?

There is still a small gain (~0.2%) for this benchmark.

Thanks,
Song
>
> >
> > And none of your patches mentions the gains of this effort helping
> > with the long term advantage of centralizing the semantics for
> > permissions on memory.
>
> Another good point. Although this brings up that this interface
> "execmem" does just handle one type of permission.