[sheepdog] [PATCH] start traversing from a random node in fill_obj_list()

[Liu Yuan]

On 05/21/2012 12:11 PM, levin li wrote:

> From: levin li <xingke.lwp at taobao.com>
> 
> Every node has the same sd_node order in its epoch, so in
> fill_obj_list(), every node starts from a same node to request
> the object list, which may cause the node overload.
> 
> Indeed, we meet this problem when there's 960 nodes in our
> cluster, when in the period of fill_obj_list, some node get
> 'too many requests' in client_rx_handler(), so I change it
> to start from a random node in fill_obj_list() to make load blance.
> 
> Signed-off-by: levin li <xingke.lwp at taobao.com>
> ---
>  sheep/recovery.c |   12 +++++++++++-
>  1 file changed, 11 insertions(+), 1 deletion(-)
> 
> diff --git a/sheep/recovery.c b/sheep/recovery.c
> index e31f226..55bb122 100644
> --- a/sheep/recovery.c
> +++ b/sheep/recovery.c
> @@ -699,6 +699,8 @@ static int fill_obj_list(struct recovery_work *rw)
>  	int retry_cnt;
>  	struct sd_node *cur = rw->cur_nodes;
>  	int cur_nr = rw->cur_nr_nodes;
> +	int start = random() % cur_nr;
> +	int end = cur_nr;
>  
>  	buf = malloc(buf_size);
>  	if (!buf) {
> @@ -706,7 +708,9 @@ static int fill_obj_list(struct recovery_work *rw)
>  		rw->retry = 1;
>  		return -1;
>  	}
> -	for (i = 0; i < cur_nr; i++) {
> +
> +again:
> +	for (i = start; i < end; i++) {
>  		int buf_nr;
>  		struct sd_node *node = cur + i;
>  
> @@ -738,6 +742,12 @@ static int fill_obj_list(struct recovery_work *rw)
>  			rw->count = merge_objlist(rw->oids, rw->count, (uint64_t *)buf, buf_nr);
>  	}
>  
> +	if (start != 0 && !next_rw) {
> +		end = start;
> +		start = 0;
> +		goto again;
> +	}
> +
>  	dprintf("%d\n", rw->count);
>  	free(buf);
>  	return 0;


Applied, thanks.

Yuan