[dpdk-dev] [PATCH] sched: fix port time rounding error

[Thomas Monjalon]

Jasvinder, what is the conclusion of this patch?

21/04/2020 10:21, Dewar, Alan:
> From: Singh, Jasvinder <jasvinder.singh at intel.com> 
> > > > From: Alan Dewar <alan.dewar at att.com>
> > > >
> > > > The QoS scheduler works off port time that is computed from the 
> > > > number of CPU cycles that have elapsed since the last time the port was
> > > > polled.   It divides the number of elapsed cycles to calculate how
> > > > many bytes can be sent, however this division can generate rounding 
> > > > errors, where some fraction of a byte sent may be lost.
> > > >
> > > > Lose enough of these fractional bytes and the QoS scheduler 
> > > > underperforms.  The problem is worse with low bandwidths.
> > > >
> > > > To compensate for this rounding error this fix doesn't advance the 
> > > > port's time_cpu_cycles by the number of cycles that have elapsed, 
> > > > but by multiplying the computed number of bytes that can be sent 
> > > > (which has been rounded down) by number of cycles per byte.
> > > > This will mean that port's time_cpu_cycles will lag behind the CPU 
> > > > cycles momentarily.  At the next poll, the lag will be taken into 
> > > > account.
> > > >
> > > > Fixes: de3cfa2c98 ("sched: initial import")
> > > >
> > > > Signed-off-by: Alan Dewar <alan.dewar at att.com>
> > > > ---
> > > >  lib/librte_sched/rte_sched.c | 12 ++++++++++--
> > > >  1 file changed, 10 insertions(+), 2 deletions(-)
> > > >
> > > > diff --git a/lib/librte_sched/rte_sched.c 
> > > > b/lib/librte_sched/rte_sched.c index c0983ddda..c656dba2d 100644
> > > > --- a/lib/librte_sched/rte_sched.c
> > > > +++ b/lib/librte_sched/rte_sched.c
> > > > @@ -222,6 +222,7 @@ struct rte_sched_port {
> > > >  	uint64_t time_cpu_bytes;      /* Current CPU time measured in bytes
> > > > */
> > > >  	uint64_t time;                /* Current NIC TX time measured in bytes */
> > > >  	struct rte_reciprocal inv_cycles_per_byte; /* CPU cycles per byte 
> > > > */
> > > > +	uint64_t cycles_per_byte;
> > > >
> > > >  	/* Grinders */
> > > >  	struct rte_mbuf **pkts_out;
> > > > @@ -852,6 +853,7 @@ rte_sched_port_config(struct
> > > rte_sched_port_params
> > > > *params)
> > > >  	cycles_per_byte = (rte_get_tsc_hz() << RTE_SCHED_TIME_SHIFT)
> > > >  		/ params->rate;
> > > >  	port->inv_cycles_per_byte = rte_reciprocal_value(cycles_per_byte);
> > > > +	port->cycles_per_byte = cycles_per_byte;
> > > >
> > > >  	/* Grinders */
> > > >  	port->pkts_out = NULL;
> > > > @@ -2673,20 +2675,26 @@ static inline void 
> > > > rte_sched_port_time_resync(struct rte_sched_port *port)  {
> > > >  	uint64_t cycles = rte_get_tsc_cycles();
> > > > -	uint64_t cycles_diff = cycles - port->time_cpu_cycles;
> > > > +	uint64_t cycles_diff;
> > > >  	uint64_t bytes_diff;
> > > >  	uint32_t i;
> > > >
> > > > +	if (cycles < port->time_cpu_cycles)
> > > > +		goto end;
> > 
> > Above check seems redundant as port->time_cpu_cycles will always be less than the current cycles due to roundoff in previous iteration.
> > 
> 
> This was to catch the condition where the cycles wraps back to zero (after 100+ years?? depending on clock speed).  
> Rather than just going to end: the conditional should at least reset port->time_cpu_cycles back to zero.
> So there would be a very temporary glitch in accuracy once every 100+ years.