Re: [Qemu-devel] [PATCH v2] tests: numa-test: use -numa memdev option instead of legacy 'mem' option

[Igor Mammedov]

On Mon, 8 Jul 2019 19:25:01 -0300
Eduardo Habkost <address@hidden> wrote:

> On Thu, Jul 04, 2019 at 09:33:22AM +0200, Igor Mammedov wrote:
> > On Wed, 3 Jul 2019 16:37:02 -0300
> > Eduardo Habkost <address@hidden> wrote:  
> > > [...]  
> > > > +    TestData data = { .ram_size = 128 };
> > > >  
> > > > -    if (strcmp(arch, "aarch64") == 0) {
> > > > -        args = "-machine virt";
> > > > +    if (!strcmp(arch, "aarch64")) {
> > > > +        data.args = "-machine virt";
> > > > +    } else if (!strcmp(arch, "ppc64")) {
> > > > +        data.ram_size = 512;    
> > > 
> > > Wouldn't it be simpler to always use "-m 512M" and the same node
> > > sizes?  
> > it would, but it might make some CI instances start failing more
> > often where it used to work before. So I picked side of being
> > conservative and not to hog memory unless it's necessary.  
> 
> I don't understand.  Why would "-m 512M" make some CI instances
> start failing more often?

well, test would consume 4 times more RAM and if running in parallel
with other tests it might hit RAM limits set on test instance
(it seems travis-ci should be fine with 7G VMs (assuming it runs on them)
but who knows where else tests are running).

Anyways if you feel it should be 512, I can respin with it default
set to 512M (it will save a couple of lines, so honestly I don't see
a good reason to consume more where it's not necessary)

As for "the same node sizes?" it depends on testcases, for some it might be
fine but not for all (current numa-test cases are mostly focused on CPU
related parts but in bios-tables tests there is memory-less node testcase,
which hints that numa-test is rather incomplete).