[ofa-general] Re: [PATCH] opensm/ftree: save lft_buf memory allocations

[Yevgeny Kliteynik]

Hi Sasha,

Sasha Khapyorsky wrote:
> Hi Yevgeny,
> 
> On 11:06 Mon 24 Nov     , Yevgeny Kliteynik wrote:
>> Sasha Khapyorsky wrote:
>>> Use OpenSM switch lft_buf directly and save memory (48k per switch) in
>>> local structures.
>> Looks good, thanks.
> 
> The only potential downside I could see here - is that this will require
> some handling if we will remove new_lft field (after #1406 and other
> debugging).

Right, I thought about it too, but I decided that removing new_buf
might be not so good. Right now we have two types of routing engines:
engines that are basing their decisions on the min_hop tables, and engines
that make their own decisions and creating min_hop tables as a by-product,
just for multicast routing.
The example of latter at this point is only fat-tree routing, but I'm sure
that more will follow. New routing for 3D mesh/torus comes to mind (not
necessarily the one that was already posted to the list). For this type
of routing you will need new_buf anyway, so instead of having it inside
of every routing (as it was with fat-tree before the unicast cache and lft
simplification), we'd better have one in osm_switch_t.

-- Yevgeny

> Sasha
>