[dpdk-dev] [PATCH v12 09/14] build: optional NUMA and cpu counts detection

[Bruce Richardson]

On Thu, Nov 19, 2020 at 01:57:59PM +0000, Juraj Linkeš wrote:
> 
> 
> > -----Original Message-----
> > From: Bruce Richardson <bruce.richardson at intel.com>
> > Sent: Thursday, November 19, 2020 1:20 PM
> > To: Juraj Linkeš <juraj.linkes at pantheon.tech>
> > Cc: Thomas Monjalon <thomas at monjalon.net>; Ruifeng.Wang at arm.com;
> > Honnappa.Nagarahalli at arm.com; Phil.Yang at arm.com;
> > vcchunga at amazon.com; Dharmik.Thakkar at arm.com; jerinjacobk at gmail.com;
> > hemant.agrawal at nxp.com; ajit.khaparde at broadcom.com;
> > ferruh.yigit at intel.com; dev at dpdk.org
> > Subject: Re: [dpdk-dev] [PATCH v12 09/14] build: optional NUMA and cpu counts
> > detection
> > 
> > On Wed, Nov 18, 2020 at 03:23:13PM +0000, Juraj Linkeš wrote:
> > >
> > >
> > > > -----Original Message-----
> > > > From: Thomas Monjalon <thomas at monjalon.net>
> > > > Sent: Wednesday, November 18, 2020 3:43 PM
> > > > To: Bruce Richardson <bruce.richardson at intel.com>; Juraj Linkeš
> > > > <juraj.linkes at pantheon.tech>
> > > > Cc: Ruifeng.Wang at arm.com; Honnappa.Nagarahalli at arm.com;
> > > > Phil.Yang at arm.com; vcchunga at amazon.com; Dharmik.Thakkar at arm.com;
> > > > jerinjacobk at gmail.com; hemant.agrawal at nxp.com;
> > > > ajit.khaparde at broadcom.com; ferruh.yigit at intel.com; dev at dpdk.org
> > > > Subject: Re: [dpdk-dev] [PATCH v12 09/14] build: optional NUMA and
> > > > cpu counts detection
> > > >
> > > > 18/11/2020 15:19, Juraj Linkeš:
> > > > > From: Thomas Monjalon <thomas at monjalon.net>
> > > > > > 16/11/2020 10:13, Bruce Richardson:
> > > > > > > On Mon, Nov 16, 2020 at 08:24:48AM +0100, Thomas Monjalon wrote:
> > > > > > > > 13/11/2020 15:31, Juraj Linkeš:
> > > > > > > > > +option('max_lcores', type: 'integer', value: 0,
> > > > > > > > > +	description: 'maximum number of cores/threads supported
> > > > > > > > > +by
> > > > EAL.
> > > > > > > > > +Set to positive integer to overwrite per-arch or
> > > > > > > > > +cross-compilation
> > > > > > defaults. Set to -1 to detect the number of cores on the build
> > > > > > machine.') option('max_numa_nodes', type: 'integer', value: 0,
> > > > > > > > > +	description: 'maximum number of NUMA nodes supported by
> > > > EAL.
> > > > > > > > > +Set to positive integer to overwrite per-arch or
> > > > > > > > > +cross-compilation defaults. Set to -1 to detect the
> > > > > > > > > +number of numa nodes on the build machine.')
> > > > > > > >
> > > > > > > > First comment: I don't like having so long description.
> > > > > > > > Second: I don't understand.
> > > > > > > >
> > > > > > > > It is said the default value is 0 so I expect it means automatic
> > detection.
> > > > > > > > But later it is said -1 is for detection. So ?
> > > > > > > >
> > > > > > > Zero is for the "per-arch or cross-compilation default". This
> > > > > > > was discussed quite a bit in previous versions and this was te
> > > > > > > best compromise we could come up with. Having a default of
> > > > > > > auto-detect is definitely not something I think we should go
> > > > > > > with - just thinking of all the build CI jobs running on
> > > > > > > 2 or 4 core VMs! However, Juraj really felt there was value in
> > > > > > > having auto-detection, so it's set as a -1 value, which I'm ok with.
> > > > > >
> > > > > > The problem is that I don't understand what 0 means.
> > > > > >
> > > > >
> > > > > There are three pieces of information which we need to convey:
> > > > > 1. The default value (0) indicates that per-arch or
> > > > > cross-compilation defaults
> > > > will be used.
> > > > > 2. Positive integer values will be used instead of these defaults.
> > > >
> > > > Where these positive values come from?
> > > >
> > >
> > > From the user - they will have the option to set it to whatever the like if they
> > don't want to use defaults.
> > >
> > > > > 3. Detected values will be used for native build when the value is -1.
> > > >
> > > > Why not detect for any native build set up with 0 (default)?
> > > >
> > >
> > > I'll let Bruce explain this, but I'll just say that we wanted to make the detection
> > the default for native builds, so we're in agreement.
> > 
> > I think most of us agree that the different understanding of the term "native
> > build", is the cause of much of the disagreements and points of dispute on this
> > thread. From my view point, the term "native" can refer to:
> > 
> > 1. what meson considers a native build, i.e. one not using a cross-file 2. a build
> > for a different machine architecture to the one on the build
> >    machine (this largely overlaps with #1, except that e.g. 32-bit build on
> >    64-bit may be considered a cross-build in this case).
> > 3. a build tailored exactly for the build machine itself i.e. both ISA, and
> >    things like core counts.
> > 4. a flag passed to the compiler to indicate the uarch level of the
> >    instruction set to be used, e.g. on x86, AVX2, AVX-512 etc., based on
> >    that of the build machine.
> > 
> > Historically, IIRC, in DPDK the "RTE_MACHINE" value was originally #4 since that
> > was it's use on x86 in the first versions of DPDK. With the move from make to
> > meson, that aspect was kept, but the meaning of #1 (I think we can ignore #2)
> > also came into play. Finally, while for x86 architecture, the idea of #4 still held,
> > for ARM use #3 is of major concern.
> > 
> > Is this a fair summary?
> > 
> > Based on this, my thinking is that the current "machine" value really needs to be
> > either renamed or split into two. We need to separate out the idea of the
> > "platform" (apologies if this is not the right term), from the "instruction
> > set"/"uarch" to make it clear what the value refers to. The default "platform"
> > value should probably be "generic", and the default "instruction set" should be
> > "default", which means it's set by the "platform" value.
> > 
> > This I believe should allow the flexibility we need, i.e. to tune to the native
> > machine (case #3) above, adjust the platform to "native", while to get behaviour
> > #4, and only just the ISA level, but keep generic in terms of other values, adjust
> > the "instruction set" value. In other words, for x86 the "machine" value as used
> > becomes the "instruction set" one, while for ARM (if I understand the
> > requirements correctly) the "machine" value becomes the "platform" one.
> > 
> > Thoughts on this?
> > 
> > /Bruce
> 
> I like where this is heading.
> Using a new option to set the platform/build type will remove all the confusion, I think.
> Then the 'machine' option will just set the machine args (and RTE_MACHINE, but that doesn't do anything as far as I can tell) and it'll work just like 'max_lcores' and 'max_numa_nodes' do - set just that one thing. But I don't like using the value 'default' to mean 'set by other option' (it's more 'ignore this' than 'default'). I like 'auto' or something similar more.

I'm ok with "auto" instead of "default". If max_lcores and max_numa_nodes
stick around as options, I'd even suggest changing them to string type so
that we can use "auto" as a default value there too.

I actually think too that to avoid any confusion with what went before, we
should remove the "machine" option entirely and replace it with the
"platform" and "ISA" (or suitable name, maybe "march_flag"??) options. I'd
like to remove any expectations of behaviour similar to what went before,
otherwise we'll get stuck in lots of compatibility issues. Better to define
new options with new - clearly defined - behaviours than try clarifying an
existing option after the fact.