[ofa-general] Re: [PATCHv2] opensm/lash: Set minimum VL for LASH to use

[Sasha Khapyorsky]

Hi Hal,

On 11:46 Mon 06 Jul     , Hal Rosenstock wrote:
> 
> diff --git a/opensm/opensm/osm_ucast_lash.c b/opensm/opensm/osm_ucast_lash.c
> index 12b5e34..3db75a0 100644
> --- a/opensm/opensm/osm_ucast_lash.c
> +++ b/opensm/opensm/osm_ucast_lash.c
> @@ -478,7 +478,7 @@ static void balance_virtual_lanes(lash_t * p_lash, unsigned lanes_needed)
>  	cdg_vertex_t ****cdg_vertex_matrix = p_lash->cdg_vertex_matrix;
>  	int *num_mst_in_lane = p_lash->num_mst_in_lane;
>  	int ***virtual_location = p_lash->virtual_location;
> -	int min_filled_lane, max_filled_lane, trials;
> +	int min_filled_lane, max_filled_lane, trials, max_vl;
>  	int old_min_filled_lane, old_max_filled_lane, new_num_min_lane,
>  	    new_num_max_lane;
>  	unsigned int i, j;
> @@ -486,9 +486,11 @@ static void balance_virtual_lanes(lash_t * p_lash, unsigned lanes_needed)
>  	int next_switch2, output_link2;
>  	int stop = 0, cycle_found;
>  	int cycle_found2;
> +	unsigned start_vl = p_lash->p_osm->subn.opt.lash_start_vl;
>  
> -	max_filled_lane = 0;
> -	min_filled_lane = lanes_needed - 1;
> +	max_filled_lane = start_vl;
> +	max_vl = lanes_needed + start_vl;
> +	min_filled_lane = max_vl - 1;
>  
>  	trials = num_mst_in_lane[max_filled_lane];
>  	if (lanes_needed == 1)
> @@ -590,7 +592,7 @@ static void balance_virtual_lanes(lash_t * p_lash, unsigned lanes_needed)
>  		new_num_min_lane = MAX_INT;
>  		new_num_max_lane = 0;
>  
> -		for (i = 0; i < lanes_needed; i++) {
> +		for (i = start_vl; i < max_vl; i++) {
>  
>  			if (num_mst_in_lane[i] < new_num_min_lane) {
>  				new_num_min_lane = num_mst_in_lane[i];
> @@ -674,11 +676,12 @@ static void free_lash_structures(lash_t * p_lash)
>  	unsigned int i, j, k;
>  	unsigned num_switches = p_lash->num_switches;
>  	osm_log_t *p_log = &p_lash->p_osm->log;
> +	unsigned start_vl = p_lash->p_osm->subn.opt.lash_start_vl;
>  
>  	OSM_LOG_ENTER(p_log);
>  
>  	/* free cdg_vertex_matrix */
> -	for (i = 0; i < p_lash->vl_min; i++) {
> +	for (i = start_vl; i < p_lash->vl_min; i++) {
>  		for (j = 0; j < num_switches; j++) {
>  			for (k = 0; k < num_switches; k++)
>  				if (p_lash->cdg_vertex_matrix[i][j][k])
> @@ -715,13 +718,14 @@ static int init_lash_structures(lash_t * p_lash)
>  	osm_log_t *p_log = &p_lash->p_osm->log;
>  	int status = 0;
>  	unsigned int i, j, k;
> +	unsigned start_vl = p_lash->p_osm->subn.opt.lash_start_vl;
>  
>  	OSM_LOG_ENTER(p_log);
>  
>  	/* initialise cdg_vertex_matrix[num_switches][num_switches][num_switches] */
>  	p_lash->cdg_vertex_matrix =
> -	    (cdg_vertex_t ****) malloc(vl_min * sizeof(cdg_vertex_t ****));
> -	for (i = 0; i < vl_min; i++) {
> +	    (cdg_vertex_t ****) calloc(vl_min, sizeof(cdg_vertex_t ****));

You don't need to use calloc() when memory initialization is not
required - malloc() is always faster.

Wouldn't it be better instead of allocating extra memory for "disabled"
(0 - lash_vl_start) VLs range and tracking 'max_vl' in many places just
to setup lash->vl_min at end of discover_network_properties() as vl_min -
lash_vl_start and to increase routing_table[i].lane value by
lash_vl_start at end of lash cycle?

Sasha