[ovs-dev] [PATCH 1/2] netflow: Do 64-bit division less often.

[Ben Pfaff]

Any comments on these two patches?  They are a month old now.

On Wed, Sep 01, 2010 at 12:46:39PM -0700, Ben Pfaff wrote:
> 64-bit division is expensive.  Usually we can avoid it entirely, as done by
> this patch.
> ---
>  ofproto/netflow.c |   19 +++++++++++--------
>  1 files changed, 11 insertions(+), 8 deletions(-)
> 
> diff --git a/ofproto/netflow.c b/ofproto/netflow.c
> index 015208a..a70b2fc 100644
> --- a/ofproto/netflow.c
> +++ b/ofproto/netflow.c
> @@ -184,21 +184,24 @@ netflow_expire(struct netflow *nf, struct netflow_flow *nf_flow,
>          return;
>      }
>  
> -    /* NetFlow v5 records are limited to 32-bit counters.  If we've
> -     * wrapped a counter, send as multiple records so we don't lose
> -     * track of any traffic.  We try to evenly distribute the packet and
> -     * byte counters, so that the bytes-per-packet lengths don't look
> -     * wonky across the records. */
> -    while (byte_delta) {
> -        int n_recs = (byte_delta + UINT32_MAX - 1) / UINT32_MAX;
> +    /* NetFlow v5 records are limited to 32-bit counters.  If we've wrapped
> +     * a counter, send as multiple records so we don't lose track of any
> +     * traffic.  We try to evenly distribute the packet and byte counters,
> +     * so that the bytes-per-packet lengths don't look wonky across the
> +     * records. */
> +    while (byte_delta > UINT32_MAX) {
> +        uint32_t n_recs = byte_delta >> 32;
>          uint32_t pkt_count = pkt_delta / n_recs;
>          uint32_t byte_count = byte_delta / n_recs;
> -        
> +
>          gen_netflow_rec(nf, nf_flow, expired, pkt_count, byte_count);
>  
>          pkt_delta -= pkt_count;
>          byte_delta -= byte_count;
>      }
> +    if (byte_delta > 0) {
> +        gen_netflow_rec(nf, nf_flow, expired, pkt_delta, byte_delta);
> +    }
>  
>      /* Update flow tracking data. */
>      nf_flow->created = 0;
> -- 
> 1.7.1
>