This patch introduces a new socket option SO_SYMMETRIC_QUEUES that can be used
to enable symmetric tx and rx queues on a socket.
This option is specifically useful for epoll based multi threaded workloads
where each thread handles packets received on a single RX queue . In this model,
we have noticed that it helps to send the packets on the same TX queue
corresponding to the queue-pair associated with the RX queue specifically when
busy poll is enabled with epoll().
Two new fields are added to struct sock_common to cache the last rx ifindex and
the rx queue in the receive path of an SKB. __netdev_pick_tx() returns the cached
rx queue when this option is enabled and the TX is happening on the same device.
Signed-off-by: Sridhar Samudrala <sridhar.samudrala@intel.com>
---
include/net/request_sock.h | 1 +
include/net/sock.h | 17 +++++++++++++++++
include/uapi/asm-generic/socket.h | 2 ++
net/core/dev.c | 8 +++++++-
net/core/sock.c | 10 ++++++++++
net/ipv4/tcp_input.c | 1 +
net/ipv4/tcp_ipv4.c | 1 +
net/ipv4/tcp_minisocks.c | 1 +
8 files changed, 40 insertions(+), 1 deletion(-)
@@ -1059,6 +1059,10 @@ int sock_setsockopt(struct socket *sock, int level, int optname,sock_valbool_flag(sk,SOCK_ZEROCOPY,valbool);break;+caseSO_SYMMETRIC_QUEUES:+sk->sk_symmetric_queues=valbool;+break;+default:ret=-ENOPROTOOPT;break;
@@ -1391,6 +1395,10 @@ int sock_getsockopt(struct socket *sock, int level, int optname,v.val=sock_flag(sk,SOCK_ZEROCOPY);break;+caseSO_SYMMETRIC_QUEUES:+v.val=sk->sk_symmetric_queues;+break;+default:/* We implement the SO_SNDLOWAT etc to not be settable*(1003.1g7).
@@ -809,6 +809,7 @@ int tcp_child_process(struct sock *parent, struct sock *child,/* record NAPI ID of child */sk_mark_napi_id(child,skb);+sk_mark_rx_queue(child,skb);tcp_segs_in(tcp_sk(child),skb);if(!sock_owned_by_user(child)){
From: Alexander Duyck <hidden> Date: 2017-09-10 01:28:20
On Thu, Aug 31, 2017 at 4:27 PM, Sridhar Samudrala
[off-list ref] wrote:
quoted hunk
This patch introduces a new socket option SO_SYMMETRIC_QUEUES that can be used
to enable symmetric tx and rx queues on a socket.
This option is specifically useful for epoll based multi threaded workloads
where each thread handles packets received on a single RX queue . In this model,
we have noticed that it helps to send the packets on the same TX queue
corresponding to the queue-pair associated with the RX queue specifically when
busy poll is enabled with epoll().
Two new fields are added to struct sock_common to cache the last rx ifindex and
the rx queue in the receive path of an SKB. __netdev_pick_tx() returns the cached
rx queue when this option is enabled and the TX is happening on the same device.
Signed-off-by: Sridhar Samudrala <sridhar.samudrala@intel.com>
---
include/net/request_sock.h | 1 +
include/net/sock.h | 17 +++++++++++++++++
include/uapi/asm-generic/socket.h | 2 ++
net/core/dev.c | 8 +++++++-
net/core/sock.c | 10 ++++++++++
net/ipv4/tcp_input.c | 1 +
net/ipv4/tcp_ipv4.c | 1 +
net/ipv4/tcp_minisocks.c | 1 +
8 files changed, 40 insertions(+), 1 deletion(-)
So one thing I am not sure about is if we should be overriding XPS. It
might make sense to instead place this after XPS so that if the root
user configures it then it applies, otherwise if the socket is
requesting symmetric queues you could fall back to that, and then
finally just use hashing as the final solution for distributing the
workload.
That way if somebody decides to reserve queues for some sort of
specific traffic like AF_PACKET then they can configure the Tx via
XPS, configure the Rx via RSS redirection table reprogramming, and
then setup a filters on the hardware to direct the traffic they want
to the queues that are running AF_PACKET.
@@ -1059,6 +1059,10 @@ int sock_setsockopt(struct socket *sock, int level, int optname,sock_valbool_flag(sk,SOCK_ZEROCOPY,valbool);break;+caseSO_SYMMETRIC_QUEUES:+sk->sk_symmetric_queues=valbool;+break;+default:ret=-ENOPROTOOPT;break;
@@ -1391,6 +1395,10 @@ int sock_getsockopt(struct socket *sock, int level, int optname,v.val=sock_flag(sk,SOCK_ZEROCOPY);break;+caseSO_SYMMETRIC_QUEUES:+v.val=sk->sk_symmetric_queues;+break;+default:/* We implement the SO_SNDLOWAT etc to not be settable*(1003.1g7).
@@ -809,6 +809,7 @@ int tcp_child_process(struct sock *parent, struct sock *child,/* record NAPI ID of child */sk_mark_napi_id(child,skb);+sk_mark_rx_queue(child,skb);tcp_segs_in(tcp_sk(child),skb);if(!sock_owned_by_user(child)){--
From: Tom Herbert <hidden> Date: 2017-09-10 05:32:18
On Thu, Aug 31, 2017 at 4:27 PM, Sridhar Samudrala
[off-list ref] wrote:
This patch introduces a new socket option SO_SYMMETRIC_QUEUES that can be used
to enable symmetric tx and rx queues on a socket.
This option is specifically useful for epoll based multi threaded workloads
where each thread handles packets received on a single RX queue . In this model,
we have noticed that it helps to send the packets on the same TX queue
corresponding to the queue-pair associated with the RX queue specifically when
busy poll is enabled with epoll().
Please provide more details, test results on exactly how this helps.
Why would this better than than optimized XPS?
Thanks,
Tom
quoted hunk
Two new fields are added to struct sock_common to cache the last rx ifindex and
the rx queue in the receive path of an SKB. __netdev_pick_tx() returns the cached
rx queue when this option is enabled and the TX is happening on the same device.
Signed-off-by: Sridhar Samudrala <sridhar.samudrala@intel.com>
---
include/net/request_sock.h | 1 +
include/net/sock.h | 17 +++++++++++++++++
include/uapi/asm-generic/socket.h | 2 ++
net/core/dev.c | 8 +++++++-
net/core/sock.c | 10 ++++++++++
net/ipv4/tcp_input.c | 1 +
net/ipv4/tcp_ipv4.c | 1 +
net/ipv4/tcp_minisocks.c | 1 +
8 files changed, 40 insertions(+), 1 deletion(-)
@@ -1059,6 +1059,10 @@ int sock_setsockopt(struct socket *sock, int level, int optname,sock_valbool_flag(sk,SOCK_ZEROCOPY,valbool);break;+caseSO_SYMMETRIC_QUEUES:+sk->sk_symmetric_queues=valbool;+break;+default:ret=-ENOPROTOOPT;break;
@@ -1391,6 +1395,10 @@ int sock_getsockopt(struct socket *sock, int level, int optname,v.val=sock_flag(sk,SOCK_ZEROCOPY);break;+caseSO_SYMMETRIC_QUEUES:+v.val=sk->sk_symmetric_queues;+break;+default:/* We implement the SO_SNDLOWAT etc to not be settable*(1003.1g7).
@@ -809,6 +809,7 @@ int tcp_child_process(struct sock *parent, struct sock *child,/* record NAPI ID of child */sk_mark_napi_id(child,skb);+sk_mark_rx_queue(child,skb);tcp_segs_in(tcp_sk(child),skb);if(!sock_owned_by_user(child)){--
From: Tom Herbert <hidden> Date: 2017-09-10 15:19:41
On Thu, Aug 31, 2017 at 4:27 PM, Sridhar Samudrala
[off-list ref] wrote:
quoted hunk
This patch introduces a new socket option SO_SYMMETRIC_QUEUES that can be used
to enable symmetric tx and rx queues on a socket.
This option is specifically useful for epoll based multi threaded workloads
where each thread handles packets received on a single RX queue . In this model,
we have noticed that it helps to send the packets on the same TX queue
corresponding to the queue-pair associated with the RX queue specifically when
busy poll is enabled with epoll().
Two new fields are added to struct sock_common to cache the last rx ifindex and
the rx queue in the receive path of an SKB. __netdev_pick_tx() returns the cached
rx queue when this option is enabled and the TX is happening on the same device.
Signed-off-by: Sridhar Samudrala <sridhar.samudrala@intel.com>
---
include/net/request_sock.h | 1 +
include/net/sock.h | 17 +++++++++++++++++
include/uapi/asm-generic/socket.h | 2 ++
net/core/dev.c | 8 +++++++-
net/core/sock.c | 10 ++++++++++
net/ipv4/tcp_input.c | 1 +
net/ipv4/tcp_ipv4.c | 1 +
net/ipv4/tcp_minisocks.c | 1 +
8 files changed, 40 insertions(+), 1 deletion(-)
This enforces that notion of queue pairs which is not universal
concept to NICs. There are many devices and instances where we
purposely avoid having a 1-1 relationship between rx and tx queues.
An alternative might be to create a rx queue to tx queue map, add the
rx queue argument to get_xps_queue, and then that function can
consider the mapping. The administrator can configure the mapping as
appropriate and can select which rx queues are subject to the mapping.
quoted hunk
if (new_index < 0)
new_index = skb_tx_hash(dev, skb);
@@ -1059,6 +1059,10 @@ int sock_setsockopt(struct socket *sock, int level, int optname,sock_valbool_flag(sk,SOCK_ZEROCOPY,valbool);break;+caseSO_SYMMETRIC_QUEUES:+sk->sk_symmetric_queues=valbool;+break;+
Allowing users control over this seems problematic to me. The intent
of packet steering is to provide good loading across the whole
systems, not just for individual applications. Exposing this control
makes that mission harder.
quoted hunk
default:
ret = -ENOPROTOOPT;
break;
@@ -1391,6 +1395,10 @@ int sock_getsockopt(struct socket *sock, int level, int optname, v.val = sock_flag(sk, SOCK_ZEROCOPY); break;+ case SO_SYMMETRIC_QUEUES:+ v.val = sk->sk_symmetric_queues;+ break;+ default: /* We implement the SO_SNDLOWAT etc to not be settable * (1003.1g 7).
@@ -2738,6 +2746,8 @@ void sock_init_data(struct socket *sock, struct sock *sk) sk->sk_max_pacing_rate = ~0U; sk->sk_pacing_rate = ~0U; sk->sk_incoming_cpu = -1;+ sk->sk_rx_ifindex = -1;+ sk->sk_rx_queue_mapping = -1; /* * Before updating sk_refcnt, we must commit prior changes to memory * (Documentation/RCU/rculist_nulls.txt for details)
This could be part of sock_rps_save_rxhash instead of new functions in
core receive path. That could be renamed sock_save_rx_info or
something like that. UDP support is also lacking with this patch so
that gets solved by a common function also.
quoted hunk
if (dst) {
if (inet_sk(sk)->rx_dst_ifindex != skb->skb_iif ||
!dst->ops->check(dst, 0)) {
@@ -809,6 +809,7 @@ int tcp_child_process(struct sock *parent, struct sock *child,/* record NAPI ID of child */sk_mark_napi_id(child,skb);+sk_mark_rx_queue(child,skb);tcp_segs_in(tcp_sk(child),skb);if(!sock_owned_by_user(child)){--
On Thu, Aug 31, 2017 at 4:27 PM, Sridhar Samudrala
[off-list ref] wrote:
quoted
This patch introduces a new socket option SO_SYMMETRIC_QUEUES that can be used
to enable symmetric tx and rx queues on a socket.
This option is specifically useful for epoll based multi threaded workloads
where each thread handles packets received on a single RX queue . In this model,
we have noticed that it helps to send the packets on the same TX queue
corresponding to the queue-pair associated with the RX queue specifically when
busy poll is enabled with epoll().
Two new fields are added to struct sock_common to cache the last rx ifindex and
the rx queue in the receive path of an SKB. __netdev_pick_tx() returns the cached
rx queue when this option is enabled and the TX is happening on the same device.
Signed-off-by: Sridhar Samudrala <sridhar.samudrala@intel.com>
---
include/net/request_sock.h | 1 +
include/net/sock.h | 17 +++++++++++++++++
include/uapi/asm-generic/socket.h | 2 ++
net/core/dev.c | 8 +++++++-
net/core/sock.c | 10 ++++++++++
net/ipv4/tcp_input.c | 1 +
net/ipv4/tcp_ipv4.c | 1 +
net/ipv4/tcp_minisocks.c | 1 +
8 files changed, 40 insertions(+), 1 deletion(-)
So one thing I am not sure about is if we should be overriding XPS. It
might make sense to instead place this after XPS so that if the root
user configures it then it applies, otherwise if the socket is
requesting symmetric queues you could fall back to that, and then
finally just use hashing as the final solution for distributing the
workload.
Isn't XPS on by default and all the devices that support XPS setup XPS maps as part of
the initialization?
Are you suggesting that the root user needs to disable XPS on the specific queues before an
application can use this option to enable symmetric queues?
That way if somebody decides to reserve queues for some sort of
specific traffic like AF_PACKET then they can configure the Tx via
XPS, configure the Rx via RSS redirection table reprogramming, and
then setup a filters on the hardware to direct the traffic they want
to the queues that are running AF_PACKET.
@@ -1059,6 +1059,10 @@ int sock_setsockopt(struct socket *sock, int level, int optname,sock_valbool_flag(sk,SOCK_ZEROCOPY,valbool);break;+caseSO_SYMMETRIC_QUEUES:+sk->sk_symmetric_queues=valbool;+break;+default:ret=-ENOPROTOOPT;break;
@@ -1391,6 +1395,10 @@ int sock_getsockopt(struct socket *sock, int level, int optname,v.val=sock_flag(sk,SOCK_ZEROCOPY);break;+caseSO_SYMMETRIC_QUEUES:+v.val=sk->sk_symmetric_queues;+break;+default:/* We implement the SO_SNDLOWAT etc to not be settable*(1003.1g7).
@@ -809,6 +809,7 @@ int tcp_child_process(struct sock *parent, struct sock *child,/* record NAPI ID of child */sk_mark_napi_id(child,skb);+sk_mark_rx_queue(child,skb);tcp_segs_in(tcp_sk(child),skb);if(!sock_owned_by_user(child)){--
On Thu, Aug 31, 2017 at 4:27 PM, Sridhar Samudrala
[off-list ref] wrote:
quoted
This patch introduces a new socket option SO_SYMMETRIC_QUEUES that can be used
to enable symmetric tx and rx queues on a socket.
This option is specifically useful for epoll based multi threaded workloads
where each thread handles packets received on a single RX queue . In this model,
we have noticed that it helps to send the packets on the same TX queue
corresponding to the queue-pair associated with the RX queue specifically when
busy poll is enabled with epoll().
Two new fields are added to struct sock_common to cache the last rx ifindex and
the rx queue in the receive path of an SKB. __netdev_pick_tx() returns the cached
rx queue when this option is enabled and the TX is happening on the same device.
Signed-off-by: Sridhar Samudrala <sridhar.samudrala@intel.com>
---
include/net/request_sock.h | 1 +
include/net/sock.h | 17 +++++++++++++++++
include/uapi/asm-generic/socket.h | 2 ++
net/core/dev.c | 8 +++++++-
net/core/sock.c | 10 ++++++++++
net/ipv4/tcp_input.c | 1 +
net/ipv4/tcp_ipv4.c | 1 +
net/ipv4/tcp_minisocks.c | 1 +
8 files changed, 40 insertions(+), 1 deletion(-)
This enforces that notion of queue pairs which is not universal
concept to NICs. There are many devices and instances where we
purposely avoid having a 1-1 relationship between rx and tx queues.
Yes. This patch assumes that TX and RX queues come in pairs.
An alternative might be to create a rx queue to tx queue map, add the
rx queue argument to get_xps_queue, and then that function can
consider the mapping. The administrator can configure the mapping as
appropriate and can select which rx queues are subject to the mapping.
This alternative looks much cleaner and doesn't require the apps to configure the
queues. Do we need to support 1 to many rx to tx queue mappings?
For our symmetric queues usecase, where a single application thread is associated with
1 queue-pair, 1-1 mapping is sufficient.
Do you see any usecase where it is useful to support 1-many mappings?
I guess i can add a sysfs entry per rx-queue to setup a tx-queue OR tx-queue-map.
quoted
if (new_index < 0)
new_index = skb_tx_hash(dev, skb);
@@ -1059,6 +1059,10 @@ int sock_setsockopt(struct socket *sock, int level, int optname,sock_valbool_flag(sk,SOCK_ZEROCOPY,valbool);break;+caseSO_SYMMETRIC_QUEUES:+sk->sk_symmetric_queues=valbool;+break;+
Allowing users control over this seems problematic to me. The intent
of packet steering is to provide good loading across the whole
systems, not just for individual applications. Exposing this control
makes that mission harder.
Sure. If we can do this on a per rxqueue basis that can be configured by the administrator,
it would be a better option.
quoted
default:
ret = -ENOPROTOOPT;
break;
@@ -1391,6 +1395,10 @@ int sock_getsockopt(struct socket *sock, int level, int optname, v.val = sock_flag(sk, SOCK_ZEROCOPY); break;+ case SO_SYMMETRIC_QUEUES:+ v.val = sk->sk_symmetric_queues;+ break;+ default: /* We implement the SO_SNDLOWAT etc to not be settable * (1003.1g 7).
@@ -2738,6 +2746,8 @@ void sock_init_data(struct socket *sock, struct sock *sk) sk->sk_max_pacing_rate = ~0U; sk->sk_pacing_rate = ~0U; sk->sk_incoming_cpu = -1;+ sk->sk_rx_ifindex = -1;+ sk->sk_rx_queue_mapping = -1; /* * Before updating sk_refcnt, we must commit prior changes to memory * (Documentation/RCU/rculist_nulls.txt for details)
This could be part of sock_rps_save_rxhash instead of new functions in
core receive path. That could be renamed sock_save_rx_info or
something like that. UDP support is also lacking with this patch so
that gets solved by a common function also.
Sure. Will look into re factoring sock_rps_save_rxhash() to save the rx queue info.
quoted
if (dst) {
if (inet_sk(sk)->rx_dst_ifindex != skb->skb_iif ||
!dst->ops->check(dst, 0)) {
@@ -809,6 +809,7 @@ int tcp_child_process(struct sock *parent, struct sock *child,/* record NAPI ID of child */sk_mark_napi_id(child,skb);+sk_mark_rx_queue(child,skb);tcp_segs_in(tcp_sk(child),skb);if(!sock_owned_by_user(child)){--
From: Alexander Duyck <hidden> Date: 2017-09-11 17:48:57
On Mon, Sep 11, 2017 at 9:48 AM, Samudrala, Sridhar
[off-list ref] wrote:
On 9/9/2017 6:28 PM, Alexander Duyck wrote:
quoted
On Thu, Aug 31, 2017 at 4:27 PM, Sridhar Samudrala
[off-list ref] wrote:
quoted
This patch introduces a new socket option SO_SYMMETRIC_QUEUES that can be
used
to enable symmetric tx and rx queues on a socket.
This option is specifically useful for epoll based multi threaded
workloads
where each thread handles packets received on a single RX queue . In this
model,
we have noticed that it helps to send the packets on the same TX queue
corresponding to the queue-pair associated with the RX queue specifically
when
busy poll is enabled with epoll().
Two new fields are added to struct sock_common to cache the last rx
ifindex and
the rx queue in the receive path of an SKB. __netdev_pick_tx() returns
the cached
rx queue when this option is enabled and the TX is happening on the same
device.
Signed-off-by: Sridhar Samudrala <sridhar.samudrala@intel.com>
---
include/net/request_sock.h | 1 +
include/net/sock.h | 17 +++++++++++++++++
include/uapi/asm-generic/socket.h | 2 ++
net/core/dev.c | 8 +++++++-
net/core/sock.c | 10 ++++++++++
net/ipv4/tcp_input.c | 1 +
net/ipv4/tcp_ipv4.c | 1 +
net/ipv4/tcp_minisocks.c | 1 +
8 files changed, 40 insertions(+), 1 deletion(-)
*msg, ...)
* @skc_node: main hash linkage for various protocol lookup tables
* @skc_nulls_node: main hash linkage for TCP/UDP/UDP-Lite protocol
* @skc_tx_queue_mapping: tx queue number for this connection
+ * @skc_rx_queue_mapping: rx queue number for this connection
+ * @skc_rx_ifindex: rx ifindex for this connection
* @skc_flags: place holder for sk_flags
* %SO_LINGER (l_onoff), %SO_BROADCAST, %SO_KEEPALIVE,
* %SO_OOBINLINE settings, %SO_TIMESTAMPING settings
* @skc_incoming_cpu: record/match cpu processing incoming packets
* @skc_refcnt: reference count
+ * @skc_symmetric_queues: symmetric tx/rx queues
*
* This is the minimal network layer representation of sockets, the
header
* for struct sock and struct inet_timewait_sock.
@@ -214,6 +218,8 @@ struct sock_common { struct hlist_nulls_node skc_nulls_node; }; int skc_tx_queue_mapping;+ int skc_rx_queue_mapping;+ int skc_rx_ifindex; union { int skc_incoming_cpu; u32 skc_rcv_wnd;
So one thing I am not sure about is if we should be overriding XPS. It
might make sense to instead place this after XPS so that if the root
user configures it then it applies, otherwise if the socket is
requesting symmetric queues you could fall back to that, and then
finally just use hashing as the final solution for distributing the
workload.
Isn't XPS on by default and all the devices that support XPS setup XPS maps
as part of
the initialization?
Are you suggesting that the root user needs to disable XPS on the specific
queues before an
application can use this option to enable symmetric queues?
XPS and this symmetric queue logic won't really play well together
since they are both attempting to do the same thing but from different
ends. Some sort of priority needs to be defined, and I would place the
request of the kernel/root user above the request of a socket and/or
application.
I guess it comes down to if we let the Tx CPU pick the queue or Rx
queue, and I would argue with an XPS configuration in place the kernel
is requesting that the Tx CPU sets the Tx queue, not the Rx/Tx hash or
Rx queue. This is similar to how we handle this for routing today as
XPS will also override the Rx to Tx mapping used there.
quoted
That way if somebody decides to reserve queues for some sort of
specific traffic like AF_PACKET then they can configure the Tx via
XPS, configure the Rx via RSS redirection table reprogramming, and
then setup a filters on the hardware to direct the traffic they want
to the queues that are running AF_PACKET.
@@ -1059,6 +1059,10 @@ int sock_setsockopt(struct socket *sock, int
level, int optname,
sock_valbool_flag(sk, SOCK_ZEROCOPY, valbool);
break;
+ case SO_SYMMETRIC_QUEUES:
+ sk->sk_symmetric_queues = valbool;
+ break;
+
default:
ret = -ENOPROTOOPT;
break;
@@ -1391,6 +1395,10 @@ int sock_getsockopt(struct socket *sock, int
level, int optname,
v.val = sock_flag(sk, SOCK_ZEROCOPY);
break;
+ case SO_SYMMETRIC_QUEUES:
+ v.val = sk->sk_symmetric_queues;
+ break;
+
default:
/* We implement the SO_SNDLOWAT etc to not be settable
* (1003.1g 7).
@@ -809,6 +809,7 @@ int tcp_child_process(struct sock *parent, struct
sock *child,
/* record NAPI ID of child */
sk_mark_napi_id(child, skb);
+ sk_mark_rx_queue(child, skb);
tcp_segs_in(tcp_sk(child), skb);
if (!sock_owned_by_user(child)) {
--
1.8.3.1
From: Tom Herbert <hidden> Date: 2017-09-11 22:07:26
On Mon, Sep 11, 2017 at 9:49 AM, Samudrala, Sridhar
[off-list ref] wrote:
On 9/10/2017 8:19 AM, Tom Herbert wrote:
quoted
On Thu, Aug 31, 2017 at 4:27 PM, Sridhar Samudrala
[off-list ref] wrote:
quoted
This patch introduces a new socket option SO_SYMMETRIC_QUEUES that can be
used
to enable symmetric tx and rx queues on a socket.
This option is specifically useful for epoll based multi threaded
workloads
where each thread handles packets received on a single RX queue . In this
model,
we have noticed that it helps to send the packets on the same TX queue
corresponding to the queue-pair associated with the RX queue specifically
when
busy poll is enabled with epoll().
Two new fields are added to struct sock_common to cache the last rx
ifindex and
the rx queue in the receive path of an SKB. __netdev_pick_tx() returns
the cached
rx queue when this option is enabled and the TX is happening on the same
device.
Signed-off-by: Sridhar Samudrala <sridhar.samudrala@intel.com>
---
include/net/request_sock.h | 1 +
include/net/sock.h | 17 +++++++++++++++++
include/uapi/asm-generic/socket.h | 2 ++
net/core/dev.c | 8 +++++++-
net/core/sock.c | 10 ++++++++++
net/ipv4/tcp_input.c | 1 +
net/ipv4/tcp_ipv4.c | 1 +
net/ipv4/tcp_minisocks.c | 1 +
8 files changed, 40 insertions(+), 1 deletion(-)
*msg, ...)
* @skc_node: main hash linkage for various protocol lookup tables
* @skc_nulls_node: main hash linkage for TCP/UDP/UDP-Lite protocol
* @skc_tx_queue_mapping: tx queue number for this connection
+ * @skc_rx_queue_mapping: rx queue number for this connection
+ * @skc_rx_ifindex: rx ifindex for this connection
* @skc_flags: place holder for sk_flags
* %SO_LINGER (l_onoff), %SO_BROADCAST, %SO_KEEPALIVE,
* %SO_OOBINLINE settings, %SO_TIMESTAMPING settings
* @skc_incoming_cpu: record/match cpu processing incoming packets
* @skc_refcnt: reference count
+ * @skc_symmetric_queues: symmetric tx/rx queues
*
* This is the minimal network layer representation of sockets, the
header
* for struct sock and struct inet_timewait_sock.
@@ -214,6 +218,8 @@ struct sock_common { struct hlist_nulls_node skc_nulls_node; }; int skc_tx_queue_mapping;+ int skc_rx_queue_mapping;+ int skc_rx_ifindex; union { int skc_incoming_cpu; u32 skc_rcv_wnd;
This enforces that notion of queue pairs which is not universal
concept to NICs. There are many devices and instances where we
purposely avoid having a 1-1 relationship between rx and tx queues.
Yes. This patch assumes that TX and RX queues come in pairs.
quoted
An alternative might be to create a rx queue to tx queue map, add the
rx queue argument to get_xps_queue, and then that function can
consider the mapping. The administrator can configure the mapping as
appropriate and can select which rx queues are subject to the mapping.
This alternative looks much cleaner and doesn't require the apps to
configure the
queues. Do we need to support 1 to many rx to tx queue mappings?
For our symmetric queues usecase, where a single application thread is
associated with
1 queue-pair, 1-1 mapping is sufficient.
Do you see any usecase where it is useful to support 1-many mappings?
I guess i can add a sysfs entry per rx-queue to setup a tx-queue OR
tx-queue-map.
There is no reason do disallow 1 to many, XPS already does that. In
fact, the mapping algorithm in XSP is pretty much what is needed where
instead of mapping a CPU to a queue set, this just maps a rx queue to
queue set. ooo handling can still be done, although it might be less
critical in this case.
Tom
From: Alexander Duyck <hidden> Date: 2017-09-11 22:50:13
On Mon, Sep 11, 2017 at 3:07 PM, Tom Herbert [off-list ref] wrote:
On Mon, Sep 11, 2017 at 9:49 AM, Samudrala, Sridhar
[off-list ref] wrote:
quoted
On 9/10/2017 8:19 AM, Tom Herbert wrote:
quoted
On Thu, Aug 31, 2017 at 4:27 PM, Sridhar Samudrala
[off-list ref] wrote:
quoted
This patch introduces a new socket option SO_SYMMETRIC_QUEUES that can be
used
to enable symmetric tx and rx queues on a socket.
This option is specifically useful for epoll based multi threaded
workloads
where each thread handles packets received on a single RX queue . In this
model,
we have noticed that it helps to send the packets on the same TX queue
corresponding to the queue-pair associated with the RX queue specifically
when
busy poll is enabled with epoll().
Two new fields are added to struct sock_common to cache the last rx
ifindex and
the rx queue in the receive path of an SKB. __netdev_pick_tx() returns
the cached
rx queue when this option is enabled and the TX is happening on the same
device.
Signed-off-by: Sridhar Samudrala <sridhar.samudrala@intel.com>
---
include/net/request_sock.h | 1 +
include/net/sock.h | 17 +++++++++++++++++
include/uapi/asm-generic/socket.h | 2 ++
net/core/dev.c | 8 +++++++-
net/core/sock.c | 10 ++++++++++
net/ipv4/tcp_input.c | 1 +
net/ipv4/tcp_ipv4.c | 1 +
net/ipv4/tcp_minisocks.c | 1 +
8 files changed, 40 insertions(+), 1 deletion(-)
*msg, ...)
* @skc_node: main hash linkage for various protocol lookup tables
* @skc_nulls_node: main hash linkage for TCP/UDP/UDP-Lite protocol
* @skc_tx_queue_mapping: tx queue number for this connection
+ * @skc_rx_queue_mapping: rx queue number for this connection
+ * @skc_rx_ifindex: rx ifindex for this connection
* @skc_flags: place holder for sk_flags
* %SO_LINGER (l_onoff), %SO_BROADCAST, %SO_KEEPALIVE,
* %SO_OOBINLINE settings, %SO_TIMESTAMPING settings
* @skc_incoming_cpu: record/match cpu processing incoming packets
* @skc_refcnt: reference count
+ * @skc_symmetric_queues: symmetric tx/rx queues
*
* This is the minimal network layer representation of sockets, the
header
* for struct sock and struct inet_timewait_sock.
@@ -214,6 +218,8 @@ struct sock_common { struct hlist_nulls_node skc_nulls_node; }; int skc_tx_queue_mapping;+ int skc_rx_queue_mapping;+ int skc_rx_ifindex; union { int skc_incoming_cpu; u32 skc_rcv_wnd;
This enforces that notion of queue pairs which is not universal
concept to NICs. There are many devices and instances where we
purposely avoid having a 1-1 relationship between rx and tx queues.
Yes. This patch assumes that TX and RX queues come in pairs.
quoted
An alternative might be to create a rx queue to tx queue map, add the
rx queue argument to get_xps_queue, and then that function can
consider the mapping. The administrator can configure the mapping as
appropriate and can select which rx queues are subject to the mapping.
This alternative looks much cleaner and doesn't require the apps to
configure the
queues. Do we need to support 1 to many rx to tx queue mappings?
For our symmetric queues usecase, where a single application thread is
associated with
1 queue-pair, 1-1 mapping is sufficient.
Do you see any usecase where it is useful to support 1-many mappings?
I guess i can add a sysfs entry per rx-queue to setup a tx-queue OR
tx-queue-map.
There is no reason do disallow 1 to many, XPS already does that. In
fact, the mapping algorithm in XSP is pretty much what is needed where
instead of mapping a CPU to a queue set, this just maps a rx queue to
queue set. ooo handling can still be done, although it might be less
critical in this case.
Tom
Actually I wonder if we couldn't re-purpose the queue_mapping field
that is already there in the sk_buff for this. It might provide a more
elegant way to deal with the logic for already dealing with the
recorded Rx queue at the start of __skb_tx_hash. Then if the socket
wants this it could send down the packet with the queue_mapping field
set and we would be using some the map to figure out the Rx to Tx
mapping for either routing/bridging or this socket option.
Anyway just a thought.
- Alex
From: Tom Herbert <hidden> Date: 2017-09-12 03:12:20
On Thu, Aug 31, 2017 at 4:27 PM, Sridhar Samudrala
[off-list ref] wrote:
quoted hunk
This patch introduces a new socket option SO_SYMMETRIC_QUEUES that can be used
to enable symmetric tx and rx queues on a socket.
This option is specifically useful for epoll based multi threaded workloads
where each thread handles packets received on a single RX queue . In this model,
we have noticed that it helps to send the packets on the same TX queue
corresponding to the queue-pair associated with the RX queue specifically when
busy poll is enabled with epoll().
Two new fields are added to struct sock_common to cache the last rx ifindex and
the rx queue in the receive path of an SKB. __netdev_pick_tx() returns the cached
rx queue when this option is enabled and the TX is happening on the same device.
Signed-off-by: Sridhar Samudrala <sridhar.samudrala@intel.com>
---
include/net/request_sock.h | 1 +
include/net/sock.h | 17 +++++++++++++++++
include/uapi/asm-generic/socket.h | 2 ++
net/core/dev.c | 8 +++++++-
net/core/sock.c | 10 ++++++++++
net/ipv4/tcp_input.c | 1 +
net/ipv4/tcp_ipv4.c | 1 +
net/ipv4/tcp_minisocks.c | 1 +
8 files changed, 40 insertions(+), 1 deletion(-)
Two ints in sock_common for this purpose is quite expensive and the
use case for this is limited-- even if a RX->TX queue mapping were
introduced to eliminate the queue pair assumption this still won't
help if the receive and transmit interfaces are different for the
connection. I think we really need to see some very compelling results
to be able to justify this.
Thanks,
Tom
From: Eric Dumazet <hidden> Date: 2017-09-12 03:53:59
On Mon, 2017-09-11 at 20:12 -0700, Tom Herbert wrote:
Two ints in sock_common for this purpose is quite expensive and the
use case for this is limited-- even if a RX->TX queue mapping were
introduced to eliminate the queue pair assumption this still won't
help if the receive and transmit interfaces are different for the
connection. I think we really need to see some very compelling results
to be able to justify this.
Yes, this is unreasonable cost.
XPS should really cover the case already.
On Mon, 2017-09-11 at 20:12 -0700, Tom Herbert wrote:
quoted
Two ints in sock_common for this purpose is quite expensive and the
use case for this is limited-- even if a RX->TX queue mapping were
introduced to eliminate the queue pair assumption this still won't
help if the receive and transmit interfaces are different for the
connection. I think we really need to see some very compelling results
to be able to justify this.
Will try to collect and post some perf data with symmetric queue configuration.
Yes, this is unreasonable cost.
XPS should really cover the case already.
Eric,
Can you clarify how XPS covers the RX-> TX queue mapping case?
Is it possible to configure XPS to select TX queue based on the RX queue of a flow?
IIUC, it is based on the CPU of the thread doing the transmit OR based on skb->priority to TC mapping?
It may be possible to get this effect if the the threads are pinned to a core, but if the app threads are
freely moving, i am not sure how XPS can be configured to select the TX queue based on the RX queue of a flow.
Thanks
Sridhar
From: Eric Dumazet <hidden> Date: 2017-09-12 15:47:43
On Mon, 2017-09-11 at 23:27 -0700, Samudrala, Sridhar wrote:
On 9/11/2017 8:53 PM, Eric Dumazet wrote:
quoted
On Mon, 2017-09-11 at 20:12 -0700, Tom Herbert wrote:
quoted
Two ints in sock_common for this purpose is quite expensive and the
use case for this is limited-- even if a RX->TX queue mapping were
introduced to eliminate the queue pair assumption this still won't
help if the receive and transmit interfaces are different for the
connection. I think we really need to see some very compelling results
to be able to justify this.
Will try to collect and post some perf data with symmetric queue
configuration.
quoted
Yes, this is unreasonable cost.
XPS should really cover the case already.
Eric,
Can you clarify how XPS covers the RX-> TX queue mapping case?
Is it possible to configure XPS to select TX queue based on the RX queue
of a flow?
IIUC, it is based on the CPU of the thread doing the transmit OR based
on skb->priority to TC mapping?
It may be possible to get this effect if the the threads are pinned to a
core, but if the app threads are
freely moving, i am not sure how XPS can be configured to select the TX
queue based on the RX queue of a flow.
If application is freely moving, how NIC can properly select the RX
queue so that packets are coming to the appropriate queue ?
This is called aRFS, and it does not scale to millions of flows.
We tried in the past, and this went nowhere really, since the setup cost
is prohibitive and DDOS vulnerable.
XPS will follow the thread, since selection is done on current cpu.
The problem is RX side. If application is free to migrate, then special
support (aRFS) is needed from the hardware.
At least for passive connections, we already have all the support in the
kernel so that you can have one thread per NIC queue, dealing with
sockets that have incoming packets all received on one NIC RX queue.
(And of course all TX packets will use the symmetric TX queue)
SO_REUSEPORT plus appropriate BPF filter can achieve that.
Say you have 32 queues, 32 cpus.
Simply use 32 listeners, 32 threads (or 32 pools of threads)
On Mon, 2017-09-11 at 23:27 -0700, Samudrala, Sridhar wrote:
quoted
On 9/11/2017 8:53 PM, Eric Dumazet wrote:
quoted
On Mon, 2017-09-11 at 20:12 -0700, Tom Herbert wrote:
quoted
Two ints in sock_common for this purpose is quite expensive and the
use case for this is limited-- even if a RX->TX queue mapping were
introduced to eliminate the queue pair assumption this still won't
help if the receive and transmit interfaces are different for the
connection. I think we really need to see some very compelling results
to be able to justify this.
Will try to collect and post some perf data with symmetric queue
configuration.
quoted
Yes, this is unreasonable cost.
XPS should really cover the case already.
Eric,
Can you clarify how XPS covers the RX-> TX queue mapping case?
Is it possible to configure XPS to select TX queue based on the RX queue
of a flow?
IIUC, it is based on the CPU of the thread doing the transmit OR based
on skb->priority to TC mapping?
It may be possible to get this effect if the the threads are pinned to a
core, but if the app threads are
freely moving, i am not sure how XPS can be configured to select the TX
queue based on the RX queue of a flow.
If application is freely moving, how NIC can properly select the RX
queue so that packets are coming to the appropriate queue ?
The RX queue is selected via RSS and we don't want to move the flow based on
where the thread is running.
This is called aRFS, and it does not scale to millions of flows.
We tried in the past, and this went nowhere really, since the setup cost
is prohibitive and DDOS vulnerable.
XPS will follow the thread, since selection is done on current cpu.
The problem is RX side. If application is free to migrate, then special
support (aRFS) is needed from the hardware.
This may be true if most of the rx processing is happening in the interrupt context.
But with busy polling, i think we don't need aRFS as a thread should be able to poll
any queue irrespective of where it is running.
At least for passive connections, we already have all the support in the
kernel so that you can have one thread per NIC queue, dealing with
sockets that have incoming packets all received on one NIC RX queue.
(And of course all TX packets will use the symmetric TX queue)
SO_REUSEPORT plus appropriate BPF filter can achieve that.
Say you have 32 queues, 32 cpus.
Simply use 32 listeners, 32 threads (or 32 pools of threads)
Yes. This will work if each thread is pinned to a core associated with the RX interrupt.
It may not be possible to pin the threads to a core.
Instead we want to associate a thread to a queue and do all the RX and TX completion
of a queue in the same thread context via busy polling.
Thanks
Sridhar
From: Tom Herbert <hidden> Date: 2017-09-12 22:53:13
On Tue, Sep 12, 2017 at 3:31 PM, Samudrala, Sridhar
[off-list ref] wrote:
On 9/12/2017 8:47 AM, Eric Dumazet wrote:
quoted
On Mon, 2017-09-11 at 23:27 -0700, Samudrala, Sridhar wrote:
quoted
On 9/11/2017 8:53 PM, Eric Dumazet wrote:
quoted
On Mon, 2017-09-11 at 20:12 -0700, Tom Herbert wrote:
quoted
Two ints in sock_common for this purpose is quite expensive and the
use case for this is limited-- even if a RX->TX queue mapping were
introduced to eliminate the queue pair assumption this still won't
help if the receive and transmit interfaces are different for the
connection. I think we really need to see some very compelling results
to be able to justify this.
Will try to collect and post some perf data with symmetric queue
configuration.
quoted
Yes, this is unreasonable cost.
XPS should really cover the case already.
Eric,
Can you clarify how XPS covers the RX-> TX queue mapping case?
Is it possible to configure XPS to select TX queue based on the RX queue
of a flow?
IIUC, it is based on the CPU of the thread doing the transmit OR based
on skb->priority to TC mapping?
It may be possible to get this effect if the the threads are pinned to a
core, but if the app threads are
freely moving, i am not sure how XPS can be configured to select the TX
queue based on the RX queue of a flow.
If application is freely moving, how NIC can properly select the RX
queue so that packets are coming to the appropriate queue ?
The RX queue is selected via RSS and we don't want to move the flow based on
where the thread is running.
Unless flow director is enabled on the Intel device... This was, I
believe, one of the first attempts to introduce a queue pair notion to
general purpose NICs. The idea was that the device records the TX
queue for a flow and then uses that to determine receive queue in a
symmetric fashion. aRFS is similar, but was under SW control how the
mapping is done. As Eric mentioned there are scalability issues with
these mechanisms, but we also found that flow director can easily
reorder packets whenever the thread moves.
quoted
This is called aRFS, and it does not scale to millions of flows.
We tried in the past, and this went nowhere really, since the setup cost
is prohibitive and DDOS vulnerable.
XPS will follow the thread, since selection is done on current cpu.
The problem is RX side. If application is free to migrate, then special
support (aRFS) is needed from the hardware.
This may be true if most of the rx processing is happening in the interrupt
context.
But with busy polling, i think we don't need aRFS as a thread should be
able to poll
any queue irrespective of where it is running.
It's not just a problem with interrupt processing, in general we like
to have all receive processing an subsequent transmit of a reply to be
done on one CPU. Silo'ing is good for performance and parallelism.
This can sometimes be relaxed in situations where CPUs share a cache
so crossing CPUs is not not costly.
quoted
At least for passive connections, we already have all the support in the
kernel so that you can have one thread per NIC queue, dealing with
sockets that have incoming packets all received on one NIC RX queue.
(And of course all TX packets will use the symmetric TX queue)
SO_REUSEPORT plus appropriate BPF filter can achieve that.
Say you have 32 queues, 32 cpus.
Simply use 32 listeners, 32 threads (or 32 pools of threads)
Yes. This will work if each thread is pinned to a core associated with the
RX interrupt.
It may not be possible to pin the threads to a core.
Instead we want to associate a thread to a queue and do all the RX and TX
completion
of a queue in the same thread context via busy polling.
When that happens it's possible for RX to be done on the completely
wrong CPU which we know is suboptimal. However, this shouldn't
negatively affect TX side since XPS will just use the queue
appropriate for running CPU. Like Eric said, this is really a receive
problem more than a transmit problem. Keeping them as independent
paths seems to be a good approach.
Tom
On Tue, Sep 12, 2017 at 3:31 PM, Samudrala, Sridhar
[off-list ref] wrote:
quoted
On 9/12/2017 8:47 AM, Eric Dumazet wrote:
quoted
On Mon, 2017-09-11 at 23:27 -0700, Samudrala, Sridhar wrote:
quoted
On 9/11/2017 8:53 PM, Eric Dumazet wrote:
quoted
On Mon, 2017-09-11 at 20:12 -0700, Tom Herbert wrote:
quoted
Two ints in sock_common for this purpose is quite expensive and the
use case for this is limited-- even if a RX->TX queue mapping were
introduced to eliminate the queue pair assumption this still won't
help if the receive and transmit interfaces are different for the
connection. I think we really need to see some very compelling results
to be able to justify this.
Will try to collect and post some perf data with symmetric queue
configuration.
Here is some performance data i collected with memcached workload over
ixgbe 10Gb NIC with mcblaster benchmark.
ixgbe is configured with 16 queues and rx-usecs is set to 1000 for a very low
interrupt rate.
ethtool -L p1p1 combined 16
ethtool -C p1p1 rx-usecs 1000
and busy poll is set to 1000usecs
sysctl net.core.busy_poll = 1000
16 threads 800K requests/sec
=============================
rtt(min/avg/max)usecs intr/sec contextswitch/sec
-----------------------------------------------------------------------
Default 2/182/10641 23391 61163
Symmetric Queues 2/50/6311 20457 32843
32 threads 800K requests/sec
=============================
rtt(min/avg/max)usecs intr/sec contextswitch/sec
------------------------------------------------------------------------
Default 2/162/6390 32168 69450
Symmetric Queues 2/50/3853 35044 35847
quoted
quoted
quoted
quoted
Yes, this is unreasonable cost.
XPS should really cover the case already.
Eric,
Can you clarify how XPS covers the RX-> TX queue mapping case?
Is it possible to configure XPS to select TX queue based on the RX queue
of a flow?
IIUC, it is based on the CPU of the thread doing the transmit OR based
on skb->priority to TC mapping?
It may be possible to get this effect if the the threads are pinned to a
core, but if the app threads are
freely moving, i am not sure how XPS can be configured to select the TX
queue based on the RX queue of a flow.
If application is freely moving, how NIC can properly select the RX
queue so that packets are coming to the appropriate queue ?
The RX queue is selected via RSS and we don't want to move the flow based on
where the thread is running.
Unless flow director is enabled on the Intel device... This was, I
believe, one of the first attempts to introduce a queue pair notion to
general purpose NICs. The idea was that the device records the TX
queue for a flow and then uses that to determine receive queue in a
symmetric fashion. aRFS is similar, but was under SW control how the
mapping is done. As Eric mentioned there are scalability issues with
these mechanisms, but we also found that flow director can easily
reorder packets whenever the thread moves.
You must be referring to the ATR(application targeted routing) feature on Intel
NICs wherea flow director entry is added for a flow based on TX queue used for
that flow. Instead, we would like to select the TX queue based on the RX queue
of a flow.
quoted
quoted
This is called aRFS, and it does not scale to millions of flows.
We tried in the past, and this went nowhere really, since the setup cost
is prohibitive and DDOS vulnerable.
XPS will follow the thread, since selection is done on current cpu.
The problem is RX side. If application is free to migrate, then special
support (aRFS) is needed from the hardware.
This may be true if most of the rx processing is happening in the interrupt
context.
But with busy polling, i think we don't need aRFS as a thread should be
able to poll
any queue irrespective of where it is running.
It's not just a problem with interrupt processing, in general we like
to have all receive processing an subsequent transmit of a reply to be
done on one CPU. Silo'ing is good for performance and parallelism.
This can sometimes be relaxed in situations where CPUs share a cache
so crossing CPUs is not not costly.
Yes. We would like to get this behavior even without binding the app thread to a CPU.
quoted
quoted
At least for passive connections, we already have all the support in the
kernel so that you can have one thread per NIC queue, dealing with
sockets that have incoming packets all received on one NIC RX queue.
(And of course all TX packets will use the symmetric TX queue)
SO_REUSEPORT plus appropriate BPF filter can achieve that.
Say you have 32 queues, 32 cpus.
Simply use 32 listeners, 32 threads (or 32 pools of threads)
Yes. This will work if each thread is pinned to a core associated with the
RX interrupt.
It may not be possible to pin the threads to a core.
Instead we want to associate a thread to a queue and do all the RX and TX
completion
of a queue in the same thread context via busy polling.
When that happens it's possible for RX to be done on the completely
wrong CPU which we know is suboptimal. However, this shouldn't
negatively affect TX side since XPS will just use the queue
appropriate for running CPU. Like Eric said, this is really a receive
problem more than a transmit problem. Keeping them as independent
paths seems to be a good approach.
We are noticing that when majority of packets are received via busy polling, it
should not be an issue if RX processing is handled by a thread running on a core
that is different from the core that is associated with the RX interrupt. Also, as
the TX completions on the associated TX queue are processed along with the RX
processing via busy polling, we would like the Transmits also to happen in the same
thread context.
Would appreciate any feedback or thoughts on optional configuration to enable selection
of TX queue based on the RX queue of a flow.
Thanks
Sridhar
From: Tom Herbert <hidden> Date: 2017-09-20 00:48:21
On Tue, Sep 19, 2017 at 5:34 PM, Samudrala, Sridhar
[off-list ref] wrote:
On 9/12/2017 3:53 PM, Tom Herbert wrote:
quoted
On Tue, Sep 12, 2017 at 3:31 PM, Samudrala, Sridhar
[off-list ref] wrote:
quoted
On 9/12/2017 8:47 AM, Eric Dumazet wrote:
quoted
On Mon, 2017-09-11 at 23:27 -0700, Samudrala, Sridhar wrote:
quoted
On 9/11/2017 8:53 PM, Eric Dumazet wrote:
quoted
On Mon, 2017-09-11 at 20:12 -0700, Tom Herbert wrote:
quoted
Two ints in sock_common for this purpose is quite expensive and the
use case for this is limited-- even if a RX->TX queue mapping were
introduced to eliminate the queue pair assumption this still won't
help if the receive and transmit interfaces are different for the
connection. I think we really need to see some very compelling
results
to be able to justify this.
Will try to collect and post some perf data with symmetric queue
configuration.
Here is some performance data i collected with memcached workload over
ixgbe 10Gb NIC with mcblaster benchmark.
ixgbe is configured with 16 queues and rx-usecs is set to 1000 for a very
low
interrupt rate.
ethtool -L p1p1 combined 16
ethtool -C p1p1 rx-usecs 1000
and busy poll is set to 1000usecs
sysctl net.core.busy_poll = 1000
16 threads 800K requests/sec
=============================
rtt(min/avg/max)usecs intr/sec contextswitch/sec
-----------------------------------------------------------------------
Default 2/182/10641 23391 61163
Symmetric Queues 2/50/6311 20457 32843
32 threads 800K requests/sec
=============================
rtt(min/avg/max)usecs intr/sec contextswitch/sec
------------------------------------------------------------------------
Default 2/162/6390 32168 69450
Symmetric Queues 2/50/3853 35044 35847
No idea what "Default" configuration is. Please report how xps_cpus is
being set, how many RSS queues there are, and what the mapping is
between RSS queues and CPUs and shared caches. Also, whether and
threads are pinned.
Thanks,
Tom
From: Hannes Frederic Sowa <hidden> Date: 2017-09-20 15:30:52
Sridhar Samudrala [off-list ref] writes:
This patch introduces a new socket option SO_SYMMETRIC_QUEUES that can be used
to enable symmetric tx and rx queues on a socket.
This option is specifically useful for epoll based multi threaded workloads
where each thread handles packets received on a single RX queue . In this model,
we have noticed that it helps to send the packets on the same TX queue
corresponding to the queue-pair associated with the RX queue specifically when
busy poll is enabled with epoll().
Two new fields are added to struct sock_common to cache the last rx ifindex and
the rx queue in the receive path of an SKB. __netdev_pick_tx() returns the cached
rx queue when this option is enabled and the TX is happening on the same device.
Would it help to make the rx and tx skb hashes symmetric
(skb_get_hash_symmetric) on request?