Re: [Xen-devel] [PATCH v3 8/8] tools: add tools support for Intel CAT

[Chao Peng <chao.p.peng@xxxxxxxxxxxxxxx>]

On Thu, Apr 02, 2015 at 11:15:53AM +0800, Chao Peng wrote:
> On Wed, Apr 01, 2015 at 10:23:01AM +0100, Ian Campbell wrote:
> > On Wed, 2015-04-01 at 17:06 +0800, Chao Peng wrote:
> > > > > If this one returns all sockets but not a specified socket data 
> > > > > (which I agreed)
> > > > > and not consider legacy cmt code, then I think I can make
> > > > > libxl_count_physical_sockets() private and move it to 
> > > > > libxl/libxl_psr.c.
> > > > 
> > > > What is the legacy cmt code? But otherwise I agree, yes.
> > > 
> > > In libxl/xl_cmdimpl.c, psr_cmt_show also calculates the socket count
> > > itself. If we want to refactor it with new libxl_count_physical_sockets
> > > then libxl_count_physical_sockets should be public, otherwise it can be
> > > private to libxl_psr.c only. From my side, both directions sound OK.
> > 
> > Ah, so we would want a "return a list of all sockets" variant of
> > libxl_psr_cmt_get_cache_occupancy too? I think that's fine, we need to
> > keep the old interface but we could easily add a new one, e.g.
> > libxl_psr_cmt_get_all_cache_occupancy (insert the word "sockets" if you
> > like).
> 
> The libxl_psr_cmt_get_cache_occupancy is actually not used in xl anymore
> and should be considered as obsolete. Substitute is
> libxl_psr_cmt_get_sample which is the common API for both cache
> occupancy and memory bandwidth(For the latter the routine will be called
> twice). So if we really want to do everything batch in libxl then
> perhaps we should do that for libxl_psr_cmt_get_sample.
> 
> The other candidate is libxl_psr_cmt_get_l3_cache_size, which can also
> do batch in libxl. 
> 
> While there is also drawback: If both
> libxl_psr_cmt_get_sample/libxl_psr_cmt_get_l3_cache_size do their
> batches in libxl, then libxl_count_physical_sockets will be called twice
> in libxl which may bring performance drop.
> 
> Another solution is only do batch for libxl_psr_cmt_get_l3_cache_size.
> It returns all sockets's cache size and meantime the socket number. Once
> we get socket number, other functions(Both libxl_psr_cmt_get_sample for
> CMT/MBM and libxl_psr_cat_get_l3_info/libxl_psr_cat_get_cbm for CAT)
> can be called socket by socket in xl.

Ian, does this solution sounds OK to you?
Chao
> > 
> > So both interfaces would be something like:
> >    int libxl_psr_....(ctx, domid, TYPE **list_r, int *nr);
> > 
> > And on success *list_r points to a newly allocated array and *nr is the
> > number of elements in that array. TYPE depends on which op it is, so for
> > cache_occupancy it seems a uint32_t.
> > 
> > Is the socket address space always contiguous and starting at zero? If
> > not then the array might need to contain (socket,TYPE) structs.
> > 
> > Ian.
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@xxxxxxxxxxxxx
> http://lists.xen.org/xen-devel

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel