[dpdk-dev] [PATCH] app/testpmd: fix txonly flow generation entropy

[dpdk-dev] [PATCH] app/testpmd: fix txonly flow generation entropy

[Viacheslav Ovsiienko]

The testpmd application in txonly forwarding mode has an option
to generate the packet flows by varying the destination IP address.
The patch increments the IP for each packet sent, this improves
the entropy and RSS distibution on the peer receiving size
is getting more uniform.

Fixes: 82010ef55e7c ("app/testpmd: make txonly mode generate multiple flows")
Cc: stable@dpdk.org

Signed-off-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
---
 app/test-pmd/txonly.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/app/test-pmd/txonly.c b/app/test-pmd/txonly.c
index 4b5bec4..8a1989f 100644
--- a/app/test-pmd/txonly.c
+++ b/app/test-pmd/txonly.c
@@ -153,7 +153,6 @@
 		const uint16_t vlan_tci_outer, const uint64_t ol_flags)
 {
 	struct rte_mbuf *pkt_segs[RTE_MAX_SEGS_PER_PKT];
-	uint8_t  ip_var = RTE_PER_LCORE(_ip_var);
 	struct rte_mbuf *pkt_seg;
 	uint32_t nb_segs, pkt_len;
 	uint8_t i;
@@ -193,6 +192,7 @@
 	copy_buf_to_pkt(&pkt_ip_hdr, sizeof(pkt_ip_hdr), pkt,
 			sizeof(struct rte_ether_hdr));
 	if (txonly_multi_flow) {
+		uint8_t  ip_var = RTE_PER_LCORE(_ip_var);
 		struct rte_ipv4_hdr *ip_hdr;
 		uint32_t addr;
 
@@ -208,6 +208,7 @@
 		 */
 		addr = (tx_ip_dst_addr | (ip_var++ << 8)) + rte_lcore_id();
 		ip_hdr->src_addr = rte_cpu_to_be_32(addr);
+		RTE_PER_LCORE(_ip_var) = ip_var;
 	}
 	copy_buf_to_pkt(&pkt_udp_hdr, sizeof(pkt_udp_hdr), pkt,
 			sizeof(struct rte_ether_hdr) +
@@ -315,7 +316,7 @@
 	fs->tx_packets += nb_tx;
 
 	if (txonly_multi_flow)
-		RTE_PER_LCORE(_ip_var) += nb_tx;
+		RTE_PER_LCORE(_ip_var) -= nb_pkt - nb_tx;
 
 #ifdef RTE_TEST_PMD_RECORD_BURST_STATS
 	fs->tx_burst_stats.pkt_burst_spread[nb_tx]++;
-- 
1.8.3.1

Re: [dpdk-dev] [dpdk-stable] [PATCH] app/testpmd: fix txonly flow generation entropy

[Ferruh Yigit]

On 2/9/2020 5:02 PM, Viacheslav Ovsiienko wrote:
> The testpmd application in txonly forwarding mode has an option
> to generate the packet flows by varying the destination IP address.
> The patch increments the IP for each packet sent, this improves
> the entropy and RSS distibution on the peer receiving size
> is getting more uniform.

The IP address already incremented for each packet sent [1], I can't see what
this patch adds.

[1]
http://lxr.dpdk.org/dpdk/v19.11/source/app/test-pmd/txonly.c#L208


btw, for reference the option mentioned is "--txonly-multi-flow" and it has been
added with:
commit: 82010ef55e7c ("app/testpmd: make txonly mode generate multiple flows")

> 
> Fixes: 82010ef55e7c ("app/testpmd: make txonly mode generate multiple flows")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>
> ---
>  app/test-pmd/txonly.c | 5 +++--
>  1 file changed, 3 insertions(+), 2 deletions(-)
> 
> diff --git a/app/test-pmd/txonly.c b/app/test-pmd/txonly.c
> index 4b5bec4..8a1989f 100644
> --- a/app/test-pmd/txonly.c
> +++ b/app/test-pmd/txonly.c
> @@ -153,7 +153,6 @@
>  		const uint16_t vlan_tci_outer, const uint64_t ol_flags)
>  {
>  	struct rte_mbuf *pkt_segs[RTE_MAX_SEGS_PER_PKT];
> -	uint8_t  ip_var = RTE_PER_LCORE(_ip_var);
>  	struct rte_mbuf *pkt_seg;
>  	uint32_t nb_segs, pkt_len;
>  	uint8_t i;
> @@ -193,6 +192,7 @@
>  	copy_buf_to_pkt(&pkt_ip_hdr, sizeof(pkt_ip_hdr), pkt,
>  			sizeof(struct rte_ether_hdr));
>  	if (txonly_multi_flow) {
> +		uint8_t  ip_var = RTE_PER_LCORE(_ip_var);
>  		struct rte_ipv4_hdr *ip_hdr;
>  		uint32_t addr;
>  
> @@ -208,6 +208,7 @@
>  		 */
>  		addr = (tx_ip_dst_addr | (ip_var++ << 8)) + rte_lcore_id();
>  		ip_hdr->src_addr = rte_cpu_to_be_32(addr);
> +		RTE_PER_LCORE(_ip_var) = ip_var;
>  	}
>  	copy_buf_to_pkt(&pkt_udp_hdr, sizeof(pkt_udp_hdr), pkt,
>  			sizeof(struct rte_ether_hdr) +
> @@ -315,7 +316,7 @@
>  	fs->tx_packets += nb_tx;
>  
>  	if (txonly_multi_flow)
> -		RTE_PER_LCORE(_ip_var) += nb_tx;
> +		RTE_PER_LCORE(_ip_var) -= nb_pkt - nb_tx;
>  
>  #ifdef RTE_TEST_PMD_RECORD_BURST_STATS
>  	fs->tx_burst_stats.pkt_burst_spread[nb_tx]++;
> 

Re: [dpdk-dev] [dpdk-stable] [PATCH] app/testpmd: fix txonly flow generation entropy

[Slava Ovsiienko]

Hi, Ferruh

> -----Original Message-----
> From: Ferruh Yigit <ferruh.yigit@intel.com>
> Sent: Monday, February 10, 2020 16:02
> To: Slava Ovsiienko <viacheslavo@mellanox.com>; dev@dpdk.org
> Cc: Thomas Monjalon <thomas@monjalon.net>;
> bernard.iremonger@intel.com; stable@dpdk.org
> Subject: Re: [dpdk-stable] [PATCH] app/testpmd: fix txonly flow generation
> entropy
> 
> On 2/9/2020 5:02 PM, Viacheslav Ovsiienko wrote:
> > The testpmd application in txonly forwarding mode has an option to
> > generate the packet flows by varying the destination IP address.
> > The patch increments the IP for each packet sent, this improves the
> > entropy and RSS distibution on the peer receiving size is getting more
> > uniform.
> 
> The IP address already incremented for each packet sent [1], I can't see what
> this patch adds.

Not exactly, that commit increments the ip_var local variable only.
On the next call ip_var is assigned from thread local storage variable
ip_var = RTE_PER_LCORE(_ip_var) - and increment is lost, it is supposed to be a bug.
We have the bursts of packets with the same IPs.

My patch stores the incremented value in the local storage  variable.  It is double checked,
after applying the patch we see the more fair RSS distribution on receiving peer.

With best regards, Slava

> 
> [1]
> https://eur03.safelinks.protection.outlook.com/?url=http%3A%2F%2Flxr.dpdk.
> org%2Fdpdk%2Fv19.11%2Fsource%2Fapp%2Ftest-
> pmd%2Ftxonly.c%23L208&amp;data=02%7C01%7Cviacheslavo%40mellanox.co
> m%7C6a5be2d6ecdd4c487e7908d7ae31dc95%7Ca652971c7d2e4d9ba6a4d149
> 256f461b%7C0%7C0%7C637169401491712071&amp;sdata=QMXvKlSVy3vAUMz
> Fzj%2ByRKKWUcVToR3hlX8S3a7V21I%3D&amp;reserved=0
> 
> 
> btw, for reference the option mentioned is "--txonly-multi-flow" and it has
> been added with:
> commit: 82010ef55e7c ("app/testpmd: make txonly mode generate multiple
> flows")
> 
Mmm, yes. My patch refers exactly to the this. Sorry, I do not understand your comment,
could you, please, clarify what do you mean?

With best regards, Slava

Re: [dpdk-dev] [dpdk-stable] [PATCH] app/testpmd: fix txonly flow generation entropy

[Ferruh Yigit]

On 2/11/2020 7:57 PM, Slava Ovsiienko wrote:
> Hi, Ferruh
> 
>> -----Original Message-----
>> From: Ferruh Yigit <ferruh.yigit@intel.com>
>> Sent: Monday, February 10, 2020 16:02
>> To: Slava Ovsiienko <viacheslavo@mellanox.com>; dev@dpdk.org
>> Cc: Thomas Monjalon <thomas@monjalon.net>;
>> bernard.iremonger@intel.com; stable@dpdk.org
>> Subject: Re: [dpdk-stable] [PATCH] app/testpmd: fix txonly flow generation
>> entropy
>>
>> On 2/9/2020 5:02 PM, Viacheslav Ovsiienko wrote:
>>> The testpmd application in txonly forwarding mode has an option to
>>> generate the packet flows by varying the destination IP address.
>>> The patch increments the IP for each packet sent, this improves the
>>> entropy and RSS distibution on the peer receiving size is getting more
>>> uniform.
>>
>> The IP address already incremented for each packet sent [1], I can't see what
>> this patch adds.
> 
> Not exactly, that commit increments the ip_var local variable only.
> On the next call ip_var is assigned from thread local storage variable
> ip_var = RTE_PER_LCORE(_ip_var) - and increment is lost, it is supposed to be a bug.
> We have the bursts of packets with the same IPs.

Ahh, current code assumes the 'pkt_burst_prepare()' done as burst. And tries to
save the increment to thread local storage in 'pkt_burst_transmit()', but since
'pkt_burst_prepare()' is per packet, this logic doesn't work.

And it looks like that logic was correct when implemented, but broken during
refactoring, can you please update fixes line accordingly:

Fixes: 01b645dcff7f ("app/testpmd: move txonly prepare in separate function")

> 
> My patch stores the incremented value in the local storage  variable.  It is double checked,
> after applying the patch we see the more fair RSS distribution on receiving peer.
> 
> With best regards, Slava
> 
>>
>> [1]
>> https://eur03.safelinks.protection.outlook.com/?url=http%3A%2F%2Flxr.dpdk.
>> org%2Fdpdk%2Fv19.11%2Fsource%2Fapp%2Ftest-
>> pmd%2Ftxonly.c%23L208&amp;data=02%7C01%7Cviacheslavo%40mellanox.co
>> m%7C6a5be2d6ecdd4c487e7908d7ae31dc95%7Ca652971c7d2e4d9ba6a4d149
>> 256f461b%7C0%7C0%7C637169401491712071&amp;sdata=QMXvKlSVy3vAUMz
>> Fzj%2ByRKKWUcVToR3hlX8S3a7V21I%3D&amp;reserved=0
>>
>>
>> btw, for reference the option mentioned is "--txonly-multi-flow" and it has
>> been added with:
>> commit: 82010ef55e7c ("app/testpmd: make txonly mode generate multiple
>> flows")
>>
> Mmm, yes. My patch refers exactly to the this. Sorry, I do not understand your comment,
> could you, please, clarify what do you mean?

Just noting down for record.

> 
> With best regards, Slava
> 

Re: [dpdk-dev] [dpdk-stable] [PATCH] app/testpmd: fix txonly flow generation entropy

[Ferruh Yigit]

On 2/9/2020 5:02 PM, Viacheslav Ovsiienko wrote:
> The testpmd application in txonly forwarding mode has an option
> to generate the packet flows by varying the destination IP address.
> The patch increments the IP for each packet sent, this improves
> the entropy and RSS distibution on the peer receiving size
> is getting more uniform.
> 
> Fixes: 82010ef55e7c ("app/testpmd: make txonly mode generate multiple flows")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Viacheslav Ovsiienko <viacheslavo@mellanox.com>

Fixes line updated to:
Fixes: 01b645dcff7f ("app/testpmd: move txonly prepare in separate function")

Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
Applied to dpdk-next-net/master, thanks.