Skip to contentSkip to conversation

net: sched: pie: align PIE implementation with RFC 8033

PATCH In review Leslie Monis 4 participants 13 messages Feb 25, 2019 – Feb 26, 2019

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

v3 cover net: sched: pie: align PIE implementation with RFC 8033

The current implementation of the PIE queuing discipline is according to the IETF draft [http://tools.ietf.org/html/draft-pan-aqm-pie-00] and the paper [PIE: A Lightweight Control Scheme to Address the Bufferbloat Problem]. However, a lot of necessary modifications and enhancements have been proposed in RFC 8033, which have not yet been incorporated in the source code of Linux. This patch series helps in achieving the same.

Performance tests carried out using Flent [https://flent.org/]

Changes from v2 to v3:
- Used div_u64() instead of direct division after explicit type casting as recommended by David

Changes from v1 to v2:
- Excluded the patch setting PIE dynamically active/inactive as the test results were unsatisfactory
- Fixed a scaling issue when adding more auto-tuning cases which caused local variables to underflow
- Changed the long if/else chain to a loop as suggested by Stephen
- Changed the position of the accu_prob variable in the pie_vars structure as recommended by Stephen

Mohit P. Tahiliani (7):
net: sched: pie: change value of QUEUE_THRESHOLD
net: sched: pie: change default value of pie_params->target
net: sched: pie: change default value of pie_params->tupdate
net: sched: pie: change initial value of pie_vars->burst_time
net: sched: pie: add more cases to auto-tune alpha and beta
net: sched: pie: add derandomization mechanism
net: sched: pie: update references

 include/uapi/linux/pkt_sched.h |   2 +-
 net/sched/sch_pie.c            | 107 ++++++++++++++++++++-------------
 2 files changed, 66 insertions(+), 43 deletions(-)
Signature · 1 line
2.17.1
#1 lore raw

PATCH v3 1/7 net: sched: pie: change value of QUEUE_THRESHOLD

From: "Mohit P. Tahiliani" <[email protected]>

RFC 8033 recommends a value of 16384 bytes for the queue threshold.

---
 net/sched/sch_pie.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
net/sched/sch_pie.c +1 -1
index d1429371592f..7778eff6cdb7 100644
@@ -31,7 +31,7 @@
 #include <net/pkt_sched.h>
 #include <net/inet_ecn.h>
 
-#define QUEUE_THRESHOLD 10000
+#define QUEUE_THRESHOLD 16384
 #define DQCOUNT_INVALID -1
 #define MAX_PROB  0xffffffff
 #define PIE_SCALE 8
Signature · 1 line
2.17.1
#2 lore raw

PATCH v3 2/7 net: sched: pie: change default value of pie_params->target

In reply to Leslie Monis

From: "Mohit P. Tahiliani" <[email protected]>

RFC 8033 suggests a default value of 15 milliseconds for the target queue delay.

---
 net/sched/sch_pie.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
net/sched/sch_pie.c +1 -1
index 7778eff6cdb7..91af9bf19852 100644
@@ -83,7 +83,7 @@ static void pie_params_init(struct pie_params *params)
 	params->beta = 20;
 	params->tupdate = usecs_to_jiffies(30 * USEC_PER_MSEC);	/* 30 ms */
 	params->limit = 1000;	/* default of 1000 packets */
-	params->target = PSCHED_NS2TICKS(20 * NSEC_PER_MSEC);	/* 20 ms */
+	params->target = PSCHED_NS2TICKS(15 * NSEC_PER_MSEC);	/* 15 ms */
 	params->ecn = false;
 	params->bytemode = false;
 }
Signature · 1 line
2.17.1
#3 lore raw

PATCH v3 3/7 net: sched: pie: change default value of pie_params->tupdate

In reply to Leslie Monis

From: "Mohit P. Tahiliani" <[email protected]>

RFC 8033 suggests a default value of 15 milliseconds for the update interval.

---
 net/sched/sch_pie.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
net/sched/sch_pie.c +1 -1
index 91af9bf19852..702f75afc312 100644
@@ -81,7 +81,7 @@ static void pie_params_init(struct pie_params *params)
 {
 	params->alpha = 2;
 	params->beta = 20;
-	params->tupdate = usecs_to_jiffies(30 * USEC_PER_MSEC);	/* 30 ms */
+	params->tupdate = usecs_to_jiffies(15 * USEC_PER_MSEC);	/* 15 ms */
 	params->limit = 1000;	/* default of 1000 packets */
 	params->target = PSCHED_NS2TICKS(15 * NSEC_PER_MSEC);	/* 15 ms */
 	params->ecn = false;
Signature · 1 line
2.17.1
#4 lore raw

PATCH v3 4/7 net: sched: pie: change initial value of pie_vars->burst_time

In reply to Leslie Monis

From: "Mohit P. Tahiliani" <[email protected]>

RFC 8033 suggests an initial value of 150 milliseconds for the maximum time allowed for a burst of packets.

---
 net/sched/sch_pie.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
net/sched/sch_pie.c +2 -2
index 702f75afc312..d88ab53593b3 100644
@@ -92,8 +92,8 @@ static void pie_vars_init(struct pie_vars *vars)
 {
 	vars->dq_count = DQCOUNT_INVALID;
 	vars->avg_dq_rate = 0;
-	/* default of 100 ms in pschedtime */
-	vars->burst_time = PSCHED_NS2TICKS(100 * NSEC_PER_MSEC);
+	/* default of 150 ms in pschedtime */
+	vars->burst_time = PSCHED_NS2TICKS(150 * NSEC_PER_MSEC);
 }
 
 static bool drop_early(struct Qdisc *sch, u32 packet_size)
Signature · 1 line
2.17.1
#5 lore raw

PATCH v3 5/7 net: sched: pie: add more cases to auto-tune alpha and beta

In reply to Leslie Monis

From: "Mohit P. Tahiliani" <[email protected]>

The current implementation scales the local alpha and beta variables in the calculate_probability function by the same amount for all values of drop probability below 1%.

RFC 8033 suggests using additional cases for auto-tuning alpha and beta when the drop probability is less than 1%.

In order to add more auto-tuning cases, MAX_PROB must be scaled by u64 instead of u32 to prevent underflow when scaling the local alpha and beta variables in the calculate_probability function.

---
 include/uapi/linux/pkt_sched.h |  2 +-
 net/sched/sch_pie.c            | 65 +++++++++++++++++-----------------
 2 files changed, 33 insertions(+), 34 deletions(-)
include/uapi/linux/pkt_sched.h +1 -1
index 0d18b1d1fbbc..1eb572ef3f27 100644
@@ -954,7 +954,7 @@ enum {
 #define TCA_PIE_MAX   (__TCA_PIE_MAX - 1)
 
 struct tc_pie_xstats {
-	__u32 prob;             /* current probability */
+	__u64 prob;             /* current probability */
 	__u32 delay;            /* current delay in ms */
 	__u32 avg_dq_rate;      /* current average dq_rate in bits/pie_time */
 	__u32 packets_in;       /* total number of packets enqueued */
net/sched/sch_pie.c +32 -33

112 lines not on this page · raw message

Signature · 1 line
2.17.1
#6 lore raw

PATCH v3 6/7 net: sched: pie: add derandomization mechanism

In reply to Leslie Monis

From: "Mohit P. Tahiliani" <[email protected]>

Random dropping of packets to achieve latency control may introduce outlier situations where packets are dropped too close to each other or too far from each other. This can cause the real drop percentage to temporarily deviate from the intended drop probability. In certain scenarios, such as a small number of simultaneous TCP flows, these deviations can cause significant deviations in link utilization and queuing latency.

RFC 8033 suggests using a derandomization mechanism to avoid these deviations.

---
 net/sched/sch_pie.c | 28 +++++++++++++++++++++++++++-
 1 file changed, 27 insertions(+), 1 deletion(-)
net/sched/sch_pie.c +27 -1
index 30f158582499..916b878d3491 100644
@@ -55,8 +55,10 @@ struct pie_vars {
 	psched_time_t qdelay_old;
 	u64 dq_count;		/* measured in bytes */
 	psched_time_t dq_tstamp;	/* drain rate */
+	u64 accu_prob;		/* accumulated drop probability */
 	u32 avg_dq_rate;	/* bytes per pschedtime tick,scaled */
 	u32 qlen_old;		/* in bytes */
+	u8 accu_prob_overflows;	/* overflows of accu_prob */
 };
 
 /* statistics gathering */
@@ -91,9 +93,11 @@ static void pie_params_init(struct pie_params *params)
 static void pie_vars_init(struct pie_vars *vars)
 {
 	vars->dq_count = DQCOUNT_INVALID;
+	vars->accu_prob = 0;
 	vars->avg_dq_rate = 0;
 	/* default of 150 ms in pschedtime */
 	vars->burst_time = PSCHED_NS2TICKS(150 * NSEC_PER_MSEC);
+	vars->accu_prob_overflows = 0;
 }
 
 static bool drop_early(struct Qdisc *sch, u32 packet_size)
@@ -128,9 +132,29 @@ static bool drop_early(struct Qdisc *sch, u32 packet_size)
 	else
 		local_prob = q->vars.prob;
 
+	if (local_prob == 0) {
+		q->vars.accu_prob = 0;
+		q->vars.accu_prob_overflows = 0;
+	}
+
+	if (local_prob > MAX_PROB - q->vars.accu_prob)
+		q->vars.accu_prob_overflows++;
+
+	q->vars.accu_prob += local_prob;
+
+	if (q->vars.accu_prob_overflows == 0 &&
+	    q->vars.accu_prob < (MAX_PROB / 100) * 85)
+		return false;
+	if (q->vars.accu_prob_overflows == 8 &&
+	    q->vars.accu_prob >= MAX_PROB / 2)
+		return true;
+
 	prandom_bytes(&rnd, 8);
-	if (rnd < local_prob)
+	if (rnd < local_prob) {
+		q->vars.accu_prob = 0;
+		q->vars.accu_prob_overflows = 0;
 		return true;
+	}
 
 	return false;
 }
@@ -168,6 +192,8 @@ static int pie_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 
 out:
 	q->stats.dropped++;
+	q->vars.accu_prob = 0;
+	q->vars.accu_prob_overflows = 0;
 	return qdisc_drop(skb, sch, to_free);
 }
 
Signature · 1 line
2.17.1
#7 lore raw

PATCH v3 7/7 net: sched: pie: update references

In reply to Leslie Monis

From: "Mohit P. Tahiliani" <[email protected]>

RFC 8033 replaces the IETF draft for PIE

---
 net/sched/sch_pie.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
net/sched/sch_pie.c +1 -3
index 916b878d3491..f8314a14a256 100644
@@ -17,9 +17,7 @@
  * University of Oslo, Norway.
  *
  * References:
- * IETF draft submission: http://tools.ietf.org/html/draft-pan-aqm-pie-00
- * IEEE  Conference on High Performance Switching and Routing 2013 :
- * "PIE: A * Lightweight Control Scheme to Address the Bufferbloat Problem"
+ * RFC 8033: https://tools.ietf.org/html/rfc8034
  */
 
 #include <linux/module.h>
Signature · 1 line
2.17.1
maintainer
#8 lore raw
In reply to Leslie Monis

From: Leslie Monis <[email protected]>
Date: Tue, 26 Feb 2019 00:39:54 +0530

Quoted 13 lines

The current implementation of the PIE queuing discipline is according to the IETF draft [http://tools.ietf.org/html/draft-pan-aqm-pie-00] and the paper [PIE: A Lightweight Control Scheme to Address the Bufferbloat Problem]. However, a lot of necessary modifications and enhancements have been proposed in RFC 8033, which have not yet been incorporated in the source code of Linux. This patch series helps in achieving the same.

Performance tests carried out using Flent [https://flent.org/]

Changes from v2 to v3:
- Used div_u64() instead of direct division after explicit type casting as recommended by David

Quoted 8 lines

Changes from v1 to v2:
- Excluded the patch setting PIE dynamically active/inactive as the test results were unsatisfactory
- Fixed a scaling issue when adding more auto-tuning cases which caused local variables to underflow
- Changed the long if/else chain to a loop as suggested by Stephen
- Changed the position of the accu_prob variable in the pie_vars structure as recommended by Stephen

Series applied, thanks.

#9 lore raw
In reply to Leslie Monis
Quoted 31 lines · On Tue, 26 Feb 2019 00:39:54 +0530 Leslie Monis <[email protected]> wrote:

The current implementation of the PIE queuing discipline is according to the IETF draft [http://tools.ietf.org/html/draft-pan-aqm-pie-00] and the paper [PIE: A Lightweight Control Scheme to Address the Bufferbloat Problem]. However, a lot of necessary modifications and enhancements have been proposed in RFC 8033, which have not yet been incorporated in the source code of Linux. This patch series helps in achieving the same.

Performance tests carried out using Flent [https://flent.org/]

Changes from v2 to v3:
- Used div_u64() instead of direct division after explicit type casting as recommended by David

Changes from v1 to v2:
- Excluded the patch setting PIE dynamically active/inactive as the test results were unsatisfactory
- Fixed a scaling issue when adding more auto-tuning cases which caused local variables to underflow
- Changed the long if/else chain to a loop as suggested by Stephen
- Changed the position of the accu_prob variable in the pie_vars structure as recommended by Stephen

Mohit P. Tahiliani (7):
net: sched: pie: change value of QUEUE_THRESHOLD
net: sched: pie: change default value of pie_params->target
net: sched: pie: change default value of pie_params->tupdate
net: sched: pie: change initial value of pie_vars->burst_time
net: sched: pie: add more cases to auto-tune alpha and beta
net: sched: pie: add derandomization mechanism
net: sched: pie: update references

Quoted 3 lines
 include/uapi/linux/pkt_sched.h |   2 +-
 net/sched/sch_pie.c            | 107 ++++++++++++++++++++-------------
 2 files changed, 66 insertions(+), 43 deletions(-)

Are you concerned at all that changes to default values might change expected behavior of existing users?

#10 lore raw
Quoted 35 lines · Stephen Hemminger
Quoted 31 lines · On Tue, 26 Feb 2019 00:39:54 +0530 Leslie Monis <[email protected]> wrote:

The current implementation of the PIE queuing discipline is according to the IETF draft [http://tools.ietf.org/html/draft-pan-aqm-pie-00] and the paper [PIE: A Lightweight Control Scheme to Address the Bufferbloat Problem]. However, a lot of necessary modifications and enhancements have been proposed in RFC 8033, which have not yet been incorporated in the source code of Linux. This patch series helps in achieving the same.

Performance tests carried out using Flent [https://flent.org/]

Changes from v2 to v3:
- Used div_u64() instead of direct division after explicit type casting as recommended by David

Changes from v1 to v2:
- Excluded the patch setting PIE dynamically active/inactive as the test results were unsatisfactory
- Fixed a scaling issue when adding more auto-tuning cases which caused local variables to underflow
- Changed the long if/else chain to a loop as suggested by Stephen
- Changed the position of the accu_prob variable in the pie_vars structure as recommended by Stephen

Mohit P. Tahiliani (7):
net: sched: pie: change value of QUEUE_THRESHOLD
net: sched: pie: change default value of pie_params->target
net: sched: pie: change default value of pie_params->tupdate
net: sched: pie: change initial value of pie_vars->burst_time
net: sched: pie: add more cases to auto-tune alpha and beta
net: sched: pie: add derandomization mechanism
net: sched: pie: update references

Quoted 6 lines
Quoted 3 lines
 include/uapi/linux/pkt_sched.h |   2 +-
 net/sched/sch_pie.c            | 107 ++++++++++++++++++++-------------
 2 files changed, 66 insertions(+), 43 deletions(-)

Are you concerned at all that changes to default values might change expected behavior of existing users?

There's existing users?

Most of these changes are really subtle and came out as we went along. I'd have to *really search* to find my notes from that period, and I have not evaluated the sum of these new changes, but overall they were a slight improvement from the defaults as we shipped it way back when.

I'm not being snarky - are there existing users? Every last person I know that ever evaluated pie switched to fq_codel. Cisco ended up going with another bufferbloat-fighting solution in one new device and that team dissolved.

docsis-pie, well, that's a standard. and that's a bit different from this pie, but this pie is closer to that pie. And it works ok.

I certainly hope leslie's group has done similar testing to what we did then ? If there's results I can dig up my old results and see what compares. I imagine they plan a paper. our old flent test code and data is all published, but am not willing to spend another minute of my life on a single queued aqm design unless someone writes me a very big check for it, which I'd then spend on fixing P4 to run fq-codel.

fq-pie (which is waiting in the wings behind this set of patches) is ok, as at least, the BSD version was competetive with fq_codel in most respects. I told 'em they needed to look hard at the rate estimator.

The ECN handling problem mentioned is there on all known versions of pie, but it's that it reverts to drop too soon for ecn to be as much benefit as it could be.

and then l4s, which is the subject of 3 upcoming talks at netdevconf, well... not gonna talk about it. I'm willing to give them their day in court.

sorry to come across as grumpy. should linux be rfc compliant even if the rfc has issues? damned if I know.

Signature · 4 lines
Dave Täht
CTO, TekLibre, LLC
http://www.teklibre.com
Tel: 1-831-205-9740
#11 lore raw
In reply to Stephen Hemminger
Quoted 34 lines · Stephen Hemminger
Quoted 31 lines · On Tue, 26 Feb 2019 00:39:54 +0530 Leslie Monis <[email protected]> wrote:

The current implementation of the PIE queuing discipline is according to the IETF draft [http://tools.ietf.org/html/draft-pan-aqm-pie-00] and the paper [PIE: A Lightweight Control Scheme to Address the Bufferbloat Problem]. However, a lot of necessary modifications and enhancements have been proposed in RFC 8033, which have not yet been incorporated in the source code of Linux. This patch series helps in achieving the same.

Performance tests carried out using Flent [https://flent.org/]

Changes from v2 to v3:
- Used div_u64() instead of direct division after explicit type casting as recommended by David

Changes from v1 to v2:
- Excluded the patch setting PIE dynamically active/inactive as the test results were unsatisfactory
- Fixed a scaling issue when adding more auto-tuning cases which caused local variables to underflow
- Changed the long if/else chain to a loop as suggested by Stephen
- Changed the position of the accu_prob variable in the pie_vars structure as recommended by Stephen

Mohit P. Tahiliani (7):
net: sched: pie: change value of QUEUE_THRESHOLD
net: sched: pie: change default value of pie_params->target
net: sched: pie: change default value of pie_params->tupdate
net: sched: pie: change initial value of pie_vars->burst_time
net: sched: pie: add more cases to auto-tune alpha and beta
net: sched: pie: add derandomization mechanism
net: sched: pie: update references

Quoted 6 lines
Quoted 3 lines
 include/uapi/linux/pkt_sched.h |   2 +-
 net/sched/sch_pie.c            | 107 ++++++++++++++++++++-------------
 2 files changed, 66 insertions(+), 43 deletions(-)

Are you concerned at all that changes to default values might change expected behavior of existing users?

Hi Stephen,

As Dave mentioned, the changes which we have made do not really change the behaviour of the aqm drastically. Our performance tests show that these changes improve performance without any side-effects. So existing users (if there are any) should not be negatively affected in any way.

Leslie

#12 lore raw
Quoted 42 lines · On Tue, 26 Feb 2019 13:50:46 +0530 Leslie Monis <[email protected]> wrote:
Quoted 34 lines · Stephen Hemminger
Quoted 31 lines · On Tue, 26 Feb 2019 00:39:54 +0530 Leslie Monis <[email protected]> wrote:

The current implementation of the PIE queuing discipline is according to the IETF draft [http://tools.ietf.org/html/draft-pan-aqm-pie-00] and the paper [PIE: A Lightweight Control Scheme to Address the Bufferbloat Problem]. However, a lot of necessary modifications and enhancements have been proposed in RFC 8033, which have not yet been incorporated in the source code of Linux. This patch series helps in achieving the same.

Performance tests carried out using Flent [https://flent.org/]

Changes from v2 to v3:
- Used div_u64() instead of direct division after explicit type casting as recommended by David

Changes from v1 to v2:
- Excluded the patch setting PIE dynamically active/inactive as the test results were unsatisfactory
- Fixed a scaling issue when adding more auto-tuning cases which caused local variables to underflow
- Changed the long if/else chain to a loop as suggested by Stephen
- Changed the position of the accu_prob variable in the pie_vars structure as recommended by Stephen

Mohit P. Tahiliani (7):
net: sched: pie: change value of QUEUE_THRESHOLD
net: sched: pie: change default value of pie_params->target
net: sched: pie: change default value of pie_params->tupdate
net: sched: pie: change initial value of pie_vars->burst_time
net: sched: pie: add more cases to auto-tune alpha and beta
net: sched: pie: add derandomization mechanism
net: sched: pie: update references

Quoted 6 lines
Quoted 3 lines
 include/uapi/linux/pkt_sched.h |   2 +-
 net/sched/sch_pie.c            | 107 ++++++++++++++++++++-------------
 2 files changed, 66 insertions(+), 43 deletions(-)  

Are you concerned at all that changes to default values might change expected behavior of existing users?

Quoted 6 lines

Hi Stephen,

As Dave mentioned, the changes which we have made do not really change the behaviour of the aqm drastically. Our performance tests show that these changes improve performance without any side-effects. So existing users (if there are any) should not be negatively affected in any way.

Thanks for answering. Looks good