[dpdk-dev] [PATCH] app/test: remove large lpm test head file

[Dai, Wei]

Hi, Bruce

How about your test result for this patch ?
Especially on performance of rule looking-up ?
There are also some replies for your comments as below.

Thanks
/ Wei 

> -----Original Message-----
> From: Richardson, Bruce
> Sent: Monday, September 26, 2016 6:07 PM
> To: Dai, Wei <wei.dai at intel.com>
> Cc: dev at dpdk.org
> Subject: Re: [PATCH] app/test: remove large lpm test head file
> 
> On Mon, Sep 26, 2016 at 05:37:05PM +0800, Wei Dai wrote:
> > remove the large file app/test/test_lpm_routes.h and add codes to
> > auto-generate similar large route rule talbe which keeps same depth
> > and IP class distribution as previous one in test_lpm_routes.h .
> > With the rule table auto-generated at run time, the performance of
> > looking up keep similar to that from pervious constant talbe.
> >
> > Signed-off-by: Wei Dai <wei.dai at intel.com>
> > ---
> >  app/test/test_lpm.c        |       2 +-
> >  app/test/test_lpm_perf.c   |     268 +-
> >  app/test/test_lpm_routes.h | 1076861
> > -----------------------------------------
> >  3 files changed, 266 insertions(+), 1076865 deletions(-)  delete mode
> > 100644 app/test/test_lpm_routes.h
> >
> > diff --git a/app/test/test_lpm.c b/app/test/test_lpm.c index
> > b6ad2eb..0952f52 100644
> > --- a/app/test/test_lpm.c
> > +++ b/app/test/test_lpm.c
> > @@ -35,10 +35,10 @@
> >  #include <stdint.h>
> >  #include <stdlib.h>
> >
> > +#include <rte_ip.h>
> >  #include <rte_lpm.h>
> >
> >  #include "test.h"
> > -#include "test_lpm_routes.h"
> >  #include "test_xmmt_ops.h"
> >
> >  #define TEST_LPM_ASSERT(cond) do
> {                                            \
> > diff --git a/app/test/test_lpm_perf.c b/app/test/test_lpm_perf.c index
> > 58eb415..5582ef4 100644
> > --- a/app/test/test_lpm_perf.c
> > +++ b/app/test/test_lpm_perf.c
> > @@ -34,14 +34,15 @@
> >  #include <stdio.h>
> >  #include <stdint.h>
> >  #include <stdlib.h>
> > +#include <math.h>
> >
> >  #include <rte_cycles.h>
> >  #include <rte_random.h>
> >  #include <rte_branch_prediction.h>
> >  #include <rte_lpm.h>
> > +#include <rte_ip.h>
> >
> >  #include "test.h"
> > -#include "test_lpm_routes.h"
> >  #include "test_xmmt_ops.h"
> >
> >  #define TEST_LPM_ASSERT(cond) do
> {                                            \
> > @@ -55,6 +56,265 @@
> >  #define BATCH_SIZE (1 << 12)
> >  #define BULK_SIZE 32
> >
> > +#define MAX_RULE_NUM (1200000)
> > +
> > +struct route_rule {
> > +	uint32_t ip;
> > +	uint8_t depth;
> > +};
> > +
> > +struct route_rule large_route_table[MAX_RULE_NUM];
> > +
> > +static uint32_t num_route_entries;  /* NUM_ROUTE_ENTRIES */ #define
> > +NUM_ROUTE_ENTRIES num_route_entries
> > +
> > +struct route_rule_count {
> > +	uint32_t total;
> > +	uint32_t a[RTE_LPM_MAX_DEPTH];
> > +	uint32_t b[RTE_LPM_MAX_DEPTH];
> > +	uint32_t c[RTE_LPM_MAX_DEPTH];
> > +	uint32_t left;
> > +	uint32_t abc[3*RTE_LPM_MAX_DEPTH];
> 
> Can you provide some comments explaining how you are generating the rules
> to test with. For example, explain why have you split the sets of rules into three,
> a, b, and c, and how you use each of those three sets. Perhaps also provide a
> comment alongside each member of the structure above.

a/b/c means IP address class. These class doesn't count IP address for local network.
Because the previous large route rule table was just dumped from a real router and
as to match similar performance, I design to keep similar depth and IP address coverage 
as previous constant table. All the numbers of each depth of each IP class are just
got from previous constant table. 

> 
> <snip>
> > +static void init_rule_count(void)
> > +{
> > +	uint32_t depth;
> > +	uint32_t count;
> > +
> > +	rule_count.left = 0;
> > +	count = 0;
> > +
> > +	for (depth = 1; depth <= RTE_LPM_MAX_DEPTH; depth++) {
> > +		count += rule_count.a[depth-1];
> > +		if (rule_count.a[depth-1])
> > +			rule_count.abc[rule_count.left++] = depth;
> > +	}
> > +
> > +	for (depth = 1; depth <= RTE_LPM_MAX_DEPTH; depth++) {
> > +		count += rule_count.b[depth-1];
> > +		if (rule_count.b[depth-1])
> > +			rule_count.abc[rule_count.left++] = 256 + depth;
> > +	}
> > +
> > +	for (depth = 1; depth <= RTE_LPM_MAX_DEPTH; depth++) {
> > +		count += rule_count.c[depth-1];
> > +		if (rule_count.c[depth-1])
> > +			rule_count.abc[rule_count.left++] = 512 + depth;
> > +	}
> > +	rule_count.total = count;
> > +}
> 
> Again, this needs a comment explaining what this function is doing, and
> how/why it is doing so.
I will give more comments to explain it in v2 patch.
By the way, rule_count.abc[ ] is for quicker generation of rules.

> 
> > +
> > +static void generate_random_rule_prefix(uint32_t ip_class, uint8_t
> > +depth) { #define IP_HEAD_MASK_A  0x00000000  /* 0xxx */ #define
> > +IP_HEAD_MASK_B  0x80000000  /* 10xx */ #define IP_HEAD_MASK_C
> > +0xC0000000  /* 110x */ #define IP_HEAD_BIT_NUM_A 1 #define
> > +IP_HEAD_BIT_NUM_B 2 #define IP_HEAD_BIT_NUM_C 3
> > +
> > +	uint32_t depth_1;
> > +	uint32_t class_depth;
> > +	uint32_t range;
> > +	uint32_t mask;
> > +	uint32_t step;
> > +	uint32_t start;
> > +	uint32_t fixed_bit_num;
> > +	uint32_t ip_head_mask;
> > +	uint32_t rule_num;
> > +	uint32_t k;
> > +	struct route_rule *ptr_rule;
> > +
> > +	depth_1 = depth - 1;
> > +
> > +	if (ip_class == 0) { /* IP Address class A */
> > +		fixed_bit_num = IP_HEAD_BIT_NUM_A;
> > +		ip_head_mask = IP_HEAD_MASK_A;
> > +		rule_num = rule_count.a[depth_1];
> > +	} else if (ip_class == 1) { /* IP Address class B */
> > +		fixed_bit_num = IP_HEAD_BIT_NUM_B;
> > +		ip_head_mask = IP_HEAD_MASK_B;
> > +		rule_num = rule_count.b[depth_1];
> > +	} else { /* IP Address class C */
> > +		fixed_bit_num = IP_HEAD_BIT_NUM_C;
> > +		ip_head_mask = IP_HEAD_MASK_C;
> > +		rule_num = rule_count.c[depth_1];
> > +	}
> > +
> > +	class_depth = depth - fixed_bit_num;
> > +	range = 1 << class_depth;
> > +	mask = range - 1;
> > +	if (range <= rule_num)
> > +		step = 1;
> > +	else
> > +		step = round((double)range / rule_num);
> > +
> > +	start = lrand48() & mask;
> > +	ptr_rule = &large_route_table[num_route_entries];
> > +	for (k = 0; k < rule_num; k++) {
> > +		ptr_rule->ip = (start << (RTE_LPM_MAX_DEPTH - depth))
> > +			| ip_head_mask;
> > +		ptr_rule->depth = depth;
> > +		ptr_rule++;
> > +		start = (start + step) & mask;
> > +	}
> > +	num_route_entries += rule_num;
> > +}
> 
> Again, comment explaining function, please.
V2 patch will give more annotations.
> 
> > +
> > +static void insert_rule_in_random_pos(uint32_t ip, uint8_t depth) {
> > +	uint32_t pos;
> > +	int try_count = 0;
> > +	struct route_rule tmp;
> > +
> > +	do {
> > +		pos = lrand48();
> > +		try_count++;
> > +	} while ((try_count < 10) && (pos > num_route_entries));
> > +
> > +	if ((pos > num_route_entries) || (pos >= MAX_RULE_NUM))
> > +		pos = num_route_entries >> 1;
> > +
> > +	tmp = large_route_table[pos];
> > +	large_route_table[pos].ip = ip;
> > +	large_route_table[pos].depth = depth;
> > +	if (num_route_entries < MAX_RULE_NUM)
> > +		large_route_table[num_route_entries++] = tmp; }
> > +
> > +static void generate_large_route_rule_table(void)
> > +{
> > +	uint32_t idx;
> > +	uint32_t ip_class;
> > +	uint8_t  depth;
> > +
> > +	memset(large_route_table, 0, sizeof(large_route_table));
> > +	init_rule_count();
> > +
> > +	idx = 0;
> > +	do {
> > +		depth = (rule_count.abc[idx] & 0xFF);
> > +		ip_class = rule_count.abc[idx] >> 8;
> > +
> > +		generate_random_rule_prefix(ip_class, depth);
> > +
> > +		rule_count.left--;
> > +		idx++;
> > +	} while (rule_count.left > 0);
> > +
> > +	insert_rule_in_random_pos(IPv4(0, 0, 0, 0), 8);
> > +	insert_rule_in_random_pos(IPv4(10, 2, 23, 147), 32);
> > +	insert_rule_in_random_pos(IPv4(192, 168, 100, 10), 24);
> > +	insert_rule_in_random_pos(IPv4(192, 168, 25, 100), 24);
> > +	insert_rule_in_random_pos(IPv4(192, 168, 129, 124), 32);
> 
> Why are you inserting 5 rules at random positions at the end? Explanatory
> comment needed, thanks.
This just keeps the same local IP address class and depth as previous constant table.
Place these local IP address in random position in the rule set rather than the end of it
or adjacent 5 lines.

> 
> 
> When running the code with the new auto-generated table, the rule add time is
> 5x longer than that with the original test. Have you investigated what causes
> this, and is there something that can be done to work around it?

I have designed a debug platform to trace each loop and each branch passed through during rule adding,
In previous constant large rule table, there are many (about ~78%) reappearance of same prefix 
(same IP prefix + same depth). But auto-generated rule table has much lower reappearance. 
This much affect compute quantity in rule adding.
To work around it, I assume that we may update the algorithm to keep same reappearance rate in each depth
and each IP address class. 
It will be much more complicated one and need much more efforts to tune it.
As the performance of looking-up from current auto-generated rule sets is similar as previous one, 
I have not gone on to pursuit much better performance of rule adding/deleting.

> 
> Regards,
> /Bruce