sched: fix port time rounding error
PATCH 1/1 sched: fix port time rounding error
From: Alan Dewar <[email protected]>
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.
--- lib/librte_sched/rte_sched.c | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-)
lib/librte_sched/rte_sched.c +10 -2
@@ -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; + + cycles_diff = cycles - port->time_cpu_cycles;/* Compute elapsed time in bytes */ bytes_diff =rte_reciprocal_divide (cycles_diff <<RTE_SCHED_TIME_SHIFT , port->inv_cycles_per_byte);/* Advance port time */ - port->time_cpu_cycles = cycles; + port->time_cpu_cycles += + (bytes_diff * port->cycles_per_byte) >>RTE_SCHED_TIME_SHIFT ; port->time_cpu_bytes += bytes_diff;if (port->time < port->time_cpu_bytes) port->time = port->time_cpu_bytes; +end:/* Reset pipe loop detection */ for (i =0 ; i < port->n_subports_per_port; i++) port->subports[i]->pipe_loop =RTE_SCHED_PIPE_INVALID ;
Signature · 1 line
2.17.1
Quoted 89 lines
-----Original Message-----
From: [email protected] <[email protected]>
Sent: Thursday, April 16, 2020 9:48 AM
To: Dumitrescu, Cristian <[email protected]>
Cc: [email protected]; Alan Dewar <[email protected]>
Subject: [PATCH] sched: fix port time rounding errorFrom: Alan Dewar <[email protected]>
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_cyclesby 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'stime_cpu_cycleswill 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 <[email protected]>--- lib/librte_sched/rte_sched.c | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-)lib/librte_sched/rte_sched.c +1
-0@@ -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 */ lib/librte_sched/rte_sched.c +1
-0struct 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 */ lib/librte_sched/rte_sched.c +8
-2port->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; + + cycles_diff = cycles - port->time_cpu_cycles;/* Compute elapsed time in bytes */ bytes_diff =rte_reciprocal_divide (cycles_diff <<RTE_SCHED_TIME_SHIFT , port->inv_cycles_per_byte);/* Advance port time */ - port->time_cpu_cycles = cycles; + port->time_cpu_cycles += + (bytes_diff * port->cycles_per_byte) >>RTE_SCHED_TIME_SHIFT ; port->time_cpu_bytes += bytes_diff;if (port->time < port->time_cpu_bytes) port->time = port->time_cpu_bytes; +end:/* Reset pipe loop detection */ lib/librte_sched/rte_sched.c +0
-1for (i =0 ; i < port->n_subports_per_port; i++) port->subports[i]->pipe_loop =RTE_SCHED_PIPE_INVALID ; --
Quoted 1 line
2.17.1
Adding Jasvinder.
Quoted 76 lines
-----Original Message-----
From: Dumitrescu, Cristian <[email protected]>
Sent: Friday, April 17, 2020 10:19 PM
To: [email protected]
Cc: [email protected]; Alan Dewar <[email protected]>; Singh, Jasvinder <[email protected]>
Subject: RE: [PATCH] sched: fix port time rounding errorlib/librte_sched/rte_sched.c rte_sched_port_config(struct
Quoted 47 lines
-----Original Message-----
From: [email protected] <[email protected]>
Sent: Thursday, April 16, 2020 9:48 AM
To: Dumitrescu, Cristian <[email protected]>
Cc: [email protected]; Alan Dewar <[email protected]>
Subject: [PATCH] sched: fix port time rounding errorFrom: Alan Dewar <[email protected]>
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_cyclesby 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'stime_cpu_cycleswill 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 <[email protected]>--- lib/librte_sched/rte_sched.c | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-)(quoted diff) +0
-0
b/lib/librte_sched/rte_sched.cindexc0983ddda..c656dba2d100644lib/librte_sched/rte_sched.c +1
-0@@ -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 */ lib/librte_sched/rte_sched.c +0
-0struct rte_mbuf **pkts_out;@@ -852,6 +853,7 @@ rte_sched_port_config(struct
rte_sched_port_paramsQuoted 16 lines
lib/librte_sched/rte_sched.c +1
-0*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 */ lib/librte_sched/rte_sched.c +1
-1port->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;
Quoted 2 lines
lib/librte_sched/rte_sched.c static inline void · line 2683
Quoted 2 lines
lib/librte_sched/rte_sched.c +2
-0+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.
Quoted 23 lines
Quoted 20 lines
lib/librte_sched/rte_sched.c +4
-2+ cycles_diff = cycles - port->time_cpu_cycles;/* Compute elapsed time in bytes */ bytes_diff =rte_reciprocal_divide (cycles_diff <<RTE_SCHED_TIME_SHIFT , port->inv_cycles_per_byte);/* Advance port time */ - port->time_cpu_cycles = cycles; + port->time_cpu_cycles += + (bytes_diff * port->cycles_per_byte) >>RTE_SCHED_TIME_SHIFT ; port->time_cpu_bytes += bytes_diff;if (port->time < port->time_cpu_bytes) port->time = port->time_cpu_bytes; +end:/* Reset pipe loop detection */ for (i =0 ; i < port->n_subports_per_port; i++) port->subports[i]->pipe_loop =RTE_SCHED_PIPE_INVALID ; --Quoted 1 line
2.17.1
Adding Jasvinder.
Quoted 84 lines
-----Original Message-----
From: Singh, Jasvinder <[email protected]>
Sent: Monday, April 20, 2020 12:23 PM
To: Dumitrescu, Cristian <[email protected]>; [email protected]
Cc: [email protected]; Alan Dewar <[email protected]>
Subject: RE: [PATCH] sched: fix port time rounding errorQuoted 75 lines
-----Original Message-----
From: Dumitrescu, Cristian <[email protected]>
Sent: Friday, April 17, 2020 10:19 PM
To: [email protected]
Cc: [email protected]; Alan Dewar <[email protected]>; Singh, Jasvinder <[email protected]>
Subject: RE: [PATCH] sched: fix port time rounding errorlib/librte_sched/rte_sched.c rte_sched_port_config(struct
Quoted 48 lines
-----Original Message-----
From: [email protected] <[email protected]>
Sent: Thursday, April 16, 2020 9:48 AM
To: Dumitrescu, Cristian <[email protected]>
Cc: [email protected]; Alan Dewar <[email protected]>
Subject: [PATCH] sched: fix port time rounding errorFrom: Alan Dewar <[email protected]>
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_cyclesby 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'stime_cpu_cycleswill 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 <[email protected]>--- lib/librte_sched/rte_sched.c | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-)(quoted diff) +0
-0
b/lib/librte_sched/rte_sched.cindexc0983ddda..c656dba2d100644lib/librte_sched/rte_sched.c +1
-0@@ -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;lib/librte_sched/rte_sched.c +0
-0/* Grinders */ struct rte_mbuf **pkts_out;@@ -852,6 +853,7 @@ rte_sched_port_config(struct
rte_sched_port_paramsQuoted 16 lines
lib/librte_sched/rte_sched.c +1
-0*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 */ lib/librte_sched/rte_sched.c +1
-1port->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;
Quoted 5 lines
Quoted 2 lines
lib/librte_sched/rte_sched.c static inline void · line 2683
Quoted 2 lines
lib/librte_sched/rte_sched.c +2
-0+if (cycles < port->time_cpu_cycles) + goto end;Above check seems redundant as
port->time_cpu_cycleswill 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.
Quoted 24 lines
Quoted 24 lines
Quoted 21 lines
lib/librte_sched/rte_sched.c +4
-2+ cycles_diff = cycles - port->time_cpu_cycles;/* Compute elapsed time in bytes */ bytes_diff =rte_reciprocal_divide (cycles_diff <<RTE_SCHED_TIME_SHIFT , port->inv_cycles_per_byte);/* Advance port time */ - port->time_cpu_cycles = cycles; + port->time_cpu_cycles += + (bytes_diff * port->cycles_per_byte) >>RTE_SCHED_TIME_SHIFT ; port->time_cpu_bytes += bytes_diff;if (port->time < port->time_cpu_bytes) port->time = port->time_cpu_bytes; +end:/* Reset pipe loop detection */ for (i =0 ; i < port->n_subports_per_port; i++) port->subports[i]->pipe_loop =RTE_SCHED_PIPE_INVALID ; --Quoted 1 line
2.17.1
Adding Jasvinder.
Jasvinder, what is the conclusion of this patch?
21/04/2020 10:21, Dewar, Alan:
Quoted 68 lines
From: Singh, Jasvinder <[email protected]>
Quoted 58 lines
Quoted 58 lines
lib/librte_sched/rte_sched.c rte_sched_port_config(struct
Quoted 41 lines
From: Alan Dewar <[email protected]>
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_cyclesby 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'stime_cpu_cycleswill 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 <[email protected]>--- lib/librte_sched/rte_sched.c | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-)(quoted diff) +0
-0
b/lib/librte_sched/rte_sched.cindexc0983ddda..c656dba2d100644lib/librte_sched/rte_sched.c +1
-0@@ -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;lib/librte_sched/rte_sched.c +0
-0/* Grinders */ struct rte_mbuf **pkts_out;@@ -852,6 +853,7 @@ rte_sched_port_config(struct
rte_sched_port_paramsQuoted 16 lines
lib/librte_sched/rte_sched.c +1
-0*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 */ lib/librte_sched/rte_sched.c +1
-1port->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;Quoted 5 lines
Quoted 2 lines
lib/librte_sched/rte_sched.c static inline void · line 2683
Quoted 2 lines
lib/librte_sched/rte_sched.c +2
-0+if (cycles < port->time_cpu_cycles) + goto end;Above check seems redundant as
port->time_cpu_cycleswill 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_cyclesback to zero. So there would be a very temporary glitch in accuracy once every 100+ years.
Quoted 84 lines
-----Original Message-----
From: Thomas Monjalon <[email protected]>
Sent: Wednesday, June 24, 2020 11:50 PM
To: Singh, Jasvinder <[email protected]>
Cc: Dumitrescu, Cristian <[email protected]>; '[email protected]' <[email protected]>; [email protected]; 'Alan Dewar' <[email protected]>; Dewar, Alan <[email protected]>
Subject: Re: [dpdk-dev] [PATCH] sched: fix port time rounding errorJasvinder, what is the conclusion of this patch?
21/04/2020 10:21, Dewar, Alan:
Quoted 5 lines
From: Singh, Jasvinder <[email protected]>
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
From: Alan Dewar <[email protected]>
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
Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
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_cyclesby 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 perbyte.
Quoted 44 lines
Quoted 44 lines
Quoted 44 lines
lib/librte_sched/rte_sched.c rte_sched_port_config(struct
Quoted 26 lines
This will mean that port's
time_cpu_cycleswill 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 <[email protected]>--- lib/librte_sched/rte_sched.c | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-)(quoted diff) +0
-0
b/lib/librte_sched/rte_sched.cindexc0983ddda..c656dba2d100644lib/librte_sched/rte_sched.c +0
-0@@ -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 */
lib/librte_sched/rte_sched.c +1
-0+uint64_t cycles_per_byte;/* Grinders */ struct rte_mbuf **pkts_out;@@ -852,6 +853,7 @@ rte_sched_port_config(struct
rte_sched_port_paramsQuoted 17 lines
lib/librte_sched/rte_sched.c +1
-0*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;lib/librte_sched/rte_sched.c +1
-1/* 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;Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
lib/librte_sched/rte_sched.c static inline void · line 2683
Quoted 2 lines
lib/librte_sched/rte_sched.c +2
-0+if (cycles < port->time_cpu_cycles) + goto end;Above check seems redundant as
port->time_cpu_cycleswill always beless than the current cycles due to roundoff in previous iteration.
Quoted 2 lines
Quoted 1 line
Quoted 7 lines
Quoted 1 line
This was to catch the condition where the cycles wraps back to zero (after
100+ years?? depending on clock speed).
Quoted 3 lines
Rather than just going to end: the conditional should at least reset port-
time_cpu_cyclesback to zero.
So there would be a very temporary glitch in accuracy once every 100+years.
Alan, Could you please resubmit the patch with above change? Other than that, patch looks good to me.
Okay will do.
Quoted 95 lines · Jasvinder
Quoted 85 lines
-----Original Message-----
From: Thomas Monjalon <[email protected]>
Sent: Wednesday, June 24, 2020 11:50 PM
To: Singh, Jasvinder <[email protected]>
Cc: Dumitrescu, Cristian <[email protected]>; '[email protected]' <[email protected]>; [email protected]; 'Alan Dewar' <[email protected]>; Dewar, Alan <[email protected]>
Subject: Re: [dpdk-dev] [PATCH] sched: fix port time rounding errorJasvinder, what is the conclusion of this patch?
21/04/2020 10:21, Dewar, Alan:
Quoted 5 lines
From: Singh, Jasvinder <[email protected]>
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
From: Alan Dewar <[email protected]>
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
Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
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_cyclesby 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 perbyte.
Quoted 44 lines
Quoted 44 lines
Quoted 44 lines
lib/librte_sched/rte_sched.c rte_sched_port_config(struct
Quoted 26 lines
This will mean that port's
time_cpu_cycleswill 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 <[email protected]>--- lib/librte_sched/rte_sched.c | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-)(quoted diff) +0
-0
b/lib/librte_sched/rte_sched.cindexc0983ddda..c656dba2d100644lib/librte_sched/rte_sched.c +0
-0@@ -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 */
lib/librte_sched/rte_sched.c +1
-0+uint64_t cycles_per_byte;/* Grinders */ struct rte_mbuf **pkts_out;@@ -852,6 +853,7 @@ rte_sched_port_config(struct
rte_sched_port_paramsQuoted 17 lines
lib/librte_sched/rte_sched.c +1
-0*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;lib/librte_sched/rte_sched.c +1
-1@@ -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;Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
lib/librte_sched/rte_sched.c static inline void · line 2683
Quoted 2 lines
lib/librte_sched/rte_sched.c +2
-0+if (cycles < port->time_cpu_cycles) + goto end;Above check seems redundant as
port->time_cpu_cycleswill always beless than the current cycles due to roundoff in previous iteration.
Quoted 2 lines
Quoted 1 line
Quoted 7 lines
Quoted 1 line
This was to catch the condition where the cycles wraps back to zero (after
100+ years?? depending on clock speed).
Quoted 3 lines
Rather than just going to end: the conditional should at least reset port-
time_cpu_cyclesback to zero.
So there would be a very temporary glitch in accuracy once every 100+years.
Alan, Could you please resubmit the patch with above change? Other than that, patch looks good to me.
PATCH v2 1/1 sched: fix port time rounding error
In reply to [email protected]From: Alan Dewar <[email protected]>
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.
v2:
If the cycles value wraps (100 year+) reset the port's cpu cycle back to zero.
--- lib/librte_sched/rte_sched.c | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-)
lib/librte_sched/rte_sched.c +9 -2
@@ -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 ;
@@ -2673,16 +2675,21 @@ 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) + port->time_cpu_cycles =0 ; + + cycles_diff = cycles - port->time_cpu_cycles;/* Compute elapsed time in bytes */ bytes_diff =rte_reciprocal_divide (cycles_diff <<RTE_SCHED_TIME_SHIFT , port->inv_cycles_per_byte);/* Advance port time */ - port->time_cpu_cycles = cycles; + port->time_cpu_cycles += + (bytes_diff * port->cycles_per_byte) >>RTE_SCHED_TIME_SHIFT ; port->time_cpu_bytes += bytes_diff;if (port->time < port->time_cpu_bytes) port->time = port->time_cpu_bytes;
Signature · 1 line
2.17.1
Cristian, Jasvinder, please review.
25/06/2020 11:59, [email protected]:
Quoted 75 lines
From: Alan Dewar <[email protected]>
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_cyclesby 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'stime_cpu_cycleswill lag behind the CPU cycles momentarily. At the next poll, the lag will be taken into account.v2:
If the cycles value wraps (100 year+) reset the port's cpu cycle back to zero.Fixes de3cfa2c98 ("sched: initial import")Signed-off-by Alan Dewar <[email protected]>--- lib/librte_sched/rte_sched.c | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-)lib/librte_sched/rte_sched.c +9
-2@@ -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 ;@@ -2673,16 +2675,21 @@ 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) + port->time_cpu_cycles =0 ; + + cycles_diff = cycles - port->time_cpu_cycles;/* Compute elapsed time in bytes */ bytes_diff =rte_reciprocal_divide (cycles_diff <<RTE_SCHED_TIME_SHIFT , port->inv_cycles_per_byte);/* Advance port time */ - port->time_cpu_cycles = cycles; + port->time_cpu_cycles += + (bytes_diff * port->cycles_per_byte) >>RTE_SCHED_TIME_SHIFT ; port->time_cpu_bytes += bytes_diff;if (port->time < port->time_cpu_bytes) port->time = port->time_cpu_bytes;
Singh, Jasvinder
Acked-byQuoted 81 lines
-----Original Message-----
From: Thomas Monjalon <[email protected]>
Sent: Sunday, July 5, 2020 9:42 PM
To: Dumitrescu, Cristian <[email protected]>; Singh, Jasvinder <[email protected]>
Cc: [email protected]; Alan Dewar <[email protected]>; [email protected]
Subject: Re: [dpdk-dev] [PATCH v2] sched: fix port time rounding errorCristian, Jasvinder, please review.
25/06/2020 11:59, [email protected]:
Quoted 35 lines
From: Alan Dewar <[email protected]>
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_cyclesby 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'stime_cpu_cycleswill lag behind the CPU cycles momentarily. At the next poll, the lag will be taken into account.v2:
If the cycles value wraps (100 year+) reset the port's cpu cycle back to zero.Fixes de3cfa2c98 ("sched: initial import")Signed-off-by Alan Dewar <[email protected]>--- lib/librte_sched/rte_sched.c | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-)(quoted diff) +0
-0
b/lib/librte_sched/rte_sched.cindexc0983ddda..7c022cd61100644lib/librte_sched/rte_sched.c struct rte_sched_port { · line 222
Quoted 1 line
lib/librte_sched/rte_sched.c +0
-0@@ -222,6 +222,7 @@ struct rte_sched_port {uint64_t time_cpu_bytes;/* Current CPU time measured in bytes */
lib/librte_sched/rte_sched.c rte_sched_port_config(struct
Quoted 7 lines
lib/librte_sched/rte_sched.c +1
-0uint64_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)Quoted 11 lines
lib/librte_sched/rte_sched.c +1
-1cycles_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 ;@@ -2673,16 +2675,21 @@ 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;lib/librte_sched/rte_sched.c static inline void · line 2686
Quoted 10 lines
lib/librte_sched/rte_sched.c +5
-0+uint64_t cycles_diff;uint64_t bytes_diff;uint32_t i; +if (cycles < port->time_cpu_cycles) + port->time_cpu_cycles =0 ; + + cycles_diff = cycles - port->time_cpu_cycles;/* Compute elapsed time in bytes */ bytes_diff = rte_reciprocal_divide(cycles_diff <<
RTE_SCHED_TIME_SHIFT,
Quoted 12 lines
lib/librte_sched/rte_sched.c static inline void
Quoted 6 lines
lib/librte_sched/rte_sched.c +2
-1port->inv_cycles_per_byte);/* Advance port time */ - port->time_cpu_cycles = cycles; + port->time_cpu_cycles += + (bytes_diff * port->cycles_per_byte) >>
RTE_SCHED_TIME_SHIFT;Quoted 4 lines
lib/librte_sched/rte_sched.c +0
-0port->time_cpu_bytes += bytes_diff;if (port->time < port->time_cpu_bytes) port->time = port->time_cpu_bytes;
Thomas Monjalon
AppliedQuoted 24 lines
Quoted 24 lines
25/06/2020 11:59, [email protected]:
Quoted 23 lines
From: Alan Dewar <[email protected]>
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_cyclesby 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'stime_cpu_cycleswill lag behind the CPU cycles momentarily. At the next poll, the lag will be taken into account.v2:
If the cycles value wraps (100 year+) reset the port's cpu cycle back to zero.
Quoted 5 lines
Quoted 3 lines
Quoted 3 lines
Fixes de3cfa2c98 ("sched: initial import")Signed-off-by Alan Dewar <[email protected]>Acked-by Jasvinder Singh <[email protected]>
Applied, thanks
Hi,
Quoted 35 lines · On 25/06/2020 10:59, [email protected] wrote:
From: Alan Dewar <[email protected]>
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_cyclesby 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'stime_cpu_cycleswill lag behind the CPU cycles momentarily. At the next poll, the lag will be taken into account.v2:
If the cycles value wraps (100 year+) reset the port's cpu cycle back to zero.Fixes de3cfa2c98 ("sched: initial import")Signed-off-by Alan Dewar <[email protected]>--- lib/librte_sched/rte_sched.c | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-)lib/librte_sched/rte_sched.c +0
-0
lib/librte_sched/rte_sched.c struct rte_sched_port { · line 226
Quoted 5 lines
lib/librte_sched/rte_sched.c +1
-0@@ -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;
I was backporting this patch to 18.11. The older ABI checker complains about this structure change.
"cycles_per_byte has been added at the middle position of this structural type."
Isn't this an ABI break? Dropping from 18.11 for time being.
Quoted 35 lines
lib/librte_sched/rte_sched.c +8
-2/* 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 ;@@ -2673,16 +2675,21 @@ 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) + port->time_cpu_cycles =0 ; + + cycles_diff = cycles - port->time_cpu_cycles;/* Compute elapsed time in bytes */ bytes_diff =rte_reciprocal_divide (cycles_diff <<RTE_SCHED_TIME_SHIFT , port->inv_cycles_per_byte);/* Advance port time */ - port->time_cpu_cycles = cycles; + port->time_cpu_cycles += + (bytes_diff * port->cycles_per_byte) >>RTE_SCHED_TIME_SHIFT ; port->time_cpu_bytes += bytes_diff;if (port->time < port->time_cpu_bytes) port->time = port->time_cpu_bytes;
Quoted 47 lines · Kevin Traynor
Hi,
Quoted 35 lines · On 25/06/2020 10:59, [email protected] wrote:
From: Alan Dewar <[email protected]>
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_cyclesby 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'stime_cpu_cycleswill lag behind the CPU cycles momentarily. At the next poll, the lag will be taken into account.v2:
If the cycles value wraps (100 year+) reset the port's cpu cycle back to zero.Fixes de3cfa2c98 ("sched: initial import")Signed-off-by Alan Dewar <[email protected]>--- lib/librte_sched/rte_sched.c | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-)lib/librte_sched/rte_sched.c +0
-0lib/librte_sched/rte_sched.c struct rte_sched_port { · line 226
Quoted 5 lines
lib/librte_sched/rte_sched.c +1
-0@@ -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;I was backporting this patch to 18.11. The older ABI checker complains about this structure change.
Quoted 4 lines
"
cycles_per_bytehas been added at the middle position of this structural type."Isn't this an ABI break? Dropping from 18.11 for time being.
So it looks like rte_sched_port * is an opaque pointers as it's contents are only known to rte_sched.c and not outside. To everyone else it is an opaque data structure, so structural changes to it would not be an ABI breakage.
Quoted 37 lines
Quoted 36 lines
lib/librte_sched/rte_sched.c +8
-2/* 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 ;@@ -2673,16 +2675,21 @@ 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) + port->time_cpu_cycles =0 ; + + cycles_diff = cycles - port->time_cpu_cycles;/* Compute elapsed time in bytes */ bytes_diff =rte_reciprocal_divide (cycles_diff <<RTE_SCHED_TIME_SHIFT , port->inv_cycles_per_byte);/* Advance port time */ - port->time_cpu_cycles = cycles; + port->time_cpu_cycles += + (bytes_diff * port->cycles_per_byte) >>RTE_SCHED_TIME_SHIFT ; port->time_cpu_bytes += bytes_diff;if (port->time < port->time_cpu_bytes) port->time = port->time_cpu_bytes;
Quoted 50 lines · Ray
Quoted 47 lines · Kevin Traynor
Hi,
Quoted 35 lines · On 25/06/2020 10:59, [email protected] wrote:
From: Alan Dewar <[email protected]>
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_cyclesby 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'stime_cpu_cycleswill lag behind the CPU cycles momentarily. At the next poll, the lag will be taken into account.v2:
If the cycles value wraps (100 year+) reset the port's cpu cycle back to zero.Fixes de3cfa2c98 ("sched: initial import")Signed-off-by Alan Dewar <[email protected]>--- lib/librte_sched/rte_sched.c | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-)lib/librte_sched/rte_sched.c +0
-0lib/librte_sched/rte_sched.c struct rte_sched_port { · line 226
Quoted 5 lines
lib/librte_sched/rte_sched.c +1
-0@@ -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;I was backporting this patch to 18.11. The older ABI checker complains about this structure change.
Quoted 9 lines
Quoted 4 lines
"
cycles_per_bytehas been added at the middle position of this structural type."Isn't this an ABI break? Dropping from 18.11 for time being.
So it looks like
rte_sched_port* is an opaque pointers as it's contents are only known torte_sched.cand not outside. To everyone else it is an opaque data structure, so structural changes to it would not be an ABI breakage.
Thanks Ray, makes sense. I've included the fix in 18.11.10-rc1.
Kevin.
Quoted 38 lines
Quoted 37 lines
Quoted 36 lines
lib/librte_sched/rte_sched.c +8
-2/* 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 ;@@ -2673,16 +2675,21 @@ 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) + port->time_cpu_cycles =0 ; + + cycles_diff = cycles - port->time_cpu_cycles;/* Compute elapsed time in bytes */ bytes_diff =rte_reciprocal_divide (cycles_diff <<RTE_SCHED_TIME_SHIFT , port->inv_cycles_per_byte);/* Advance port time */ - port->time_cpu_cycles = cycles; + port->time_cpu_cycles += + (bytes_diff * port->cycles_per_byte) >>RTE_SCHED_TIME_SHIFT ; port->time_cpu_bytes += bytes_diff;if (port->time < port->time_cpu_bytes) port->time = port->time_cpu_bytes;