| Message ID | fc8b00b9978b4f956fa705badfaa138854abf919.1319595687.git.luto@amacapital.net |
|---|---|
| State | Rejected, archived |
| Delegated to: | David Miller |
| Headers | show |
On 10/25/2011 07:25 PM, Andy Lutomirski wrote: > When talking to an unfixable interactive peer that fails to set > TCP_NODELAY, disabling delayed ACKs can help mitigate the problem. > This is an evil thing to do, but if the entire network is private, > it's not that evil. > > This works around a problem with the remote *application*, so make > it a socket option instead of a sysctl or a per-route option. > > Signed-off-by: Andy Lutomirski<luto@amacapital.net> > --- > > This patch is a bit embarrassing. We talk to remote applications over > TCP that are very much interactive but don't set TCP_NODELAY. These > applications apparently cannot be fixed. As a partial workaround, if we > ACK every incoming segment, then as long as they don't transmit two > segments per rtt, we do pretty well. Embarrassing/evil indeed - is it really something to go into the kernel? If the networks where this happens are indeed truly private, can they run a private kernel? Or use an LD_PRELOAD hack to wedge-in a setsockopt(TCP_NODELAY) call into the application? Or set something like tcp_naglim_def on the application system(s)? Or have the server application make a setsockopt(TCP_MAXSEG) call before listen() to a value one byte below that of what the application is sending? Is the application actually "virtuous" in sending logically associated data in one "send" call, and simply running afoul of Nagle+DelayedACK in having multiple distinct requests outstanding at once, or is it actually quite evil in that it is sending logically associated data in separate send calls? rick jones choir preaching follows: raj@tardy:~$ cat usenet_replies/nagle_algorithm > I'm not familiar with this issue, and I'm mostly ignorant about what > tcp does below the sockets interface. Can anybody briefly explain what > "nagle" is, and how and when to turn it off? Or point me to the > appropriate manual. In broad terms, whenever an application does a send() call, the logic of the Nagle algorithm is supposed to go something like this: 1) Is the quantity of data in this send, plus any queued, unsent data, greater than the MSS (Maximum Segment Size) for this connection? If yes, send the data in the user's send now (modulo any other constraints such as receiver's advertised window and the TCP congestion window). If no, go to 2. 2) Is the connection to the remote otherwise idle? That is, is there no unACKed data outstanding on the network. If yes, send the data in the user's send now. If no, queue the data and wait. Either the application will continue to call send() with enough data to get to a full MSS-worth of data, or the remote will ACK all the currently sent, unACKed data, or our retransmission timer will expire. Now, where applications run into trouble is when they have what might be described as "write, write, read" behaviour, where they present logically associated data to the transport in separate 'send' calls and those sends are typically less than the MSS for the connection. It isn't so much that they run afoul of Nagle as they run into issues with the interaction of Nagle and the other heuristics operating on the remote. In particular, the delayed ACK heuristics. When a receiving TCP is deciding whether or not to send an ACK back to the sender, in broad handwaving terms it goes through logic similar to this: a) is there data being sent back to the sender? if yes, piggy-back the ACK on the data segment. b) is there a window update being sent back to the sender? if yes, piggy-back the ACK on the window update. c) has the standalone ACK timer expired. Window updates are generally triggered by the following heuristics: i) would the window update be for a non-trivial fraction of the window - typically somewhere at or above 1/4 the window, that is, has the application "consumed" at least that much data? if yes, send a window update. if no, check ii. ii) would the window update be for, the application "consumed," at least 2*MSS worth of data? if yes, send a window update, if no wait. Now, going back to that write, write, read application, on the sending side, the first write will be transmitted by TCP via nagle rule 2 - the connection is otherwise idle. However, the second small send will be delayed as there is at that point unACKnowledged data outstanding on the connection. At the receiver, that small TCP segment will arrive and will be passed to the application. The application does not have the entire app-level message, so it will not send a reply (data to TCP) back. The typical TCP window is much much larger than the MSS, so no window update would be triggered by heuristic i. The data just arrived and consumed by the application is < 2*MSS, so no window update from heuristic ii. Since there is no window update, no ACK is sent by heuristic b. So, that leaves heuristic c - the standalone ACK timer. That ranges anywhere between 50 and 200 milliseconds depending on the TCP stack in use. If you've read this far :) now we can take a look at the effect of various things touted as "fixes" to applications experiencing this interaction. We take as our example a client-server application where both the client and the server are implemented with a write of a small application header, followed by application data. First, the "default" case which is with Nagle enabled (TCP_NODELAY _NOT_ set) and with standard ACK behaviour: Client Server Req Header -> <- Standalone ACK after Nms Req Data -> <- Possible standalone ACK <- Rsp Header Standalone ACK -> <- Rsp Data Possible standalone ACK -> For two "messages" we end-up with at least six segments on the wire. The possible standalone ACKs will depend on whether the server's response time, or client's think time is longer than the standalone ACK interval on their respective sides. Now, if TCP_NODELAY is set we see: Client Server Req Header -> Req Data -> <- Possible Standalone ACK after Nms <- Rsp Header <- Rsp Data Possible Standalone ACK -> In theory, we are down two four segments on the wire which seems good, but frankly we can do better. First though, consider what happens when someone disables delayed ACKs Client Server Req Header -> <- Immediate Standalone ACK Req Data -> <- Immediate Standalone ACK <- Rsp Header Immediate Standalone ACK -> <- Rsp Data Immediate Standalone ACK -> Now we definitly see 8 segments on the wire. It will also be that way if both TCP_NODELAY is set and delayed ACKs are disabled. How about if the application did the "right" think in the first place? That is sent the logically associated data at the same time: Client Server Request -> <- Possible Standalone ACK <- Response Possible Standalone ACK -> We are down to two segments on the wire. For "small" packets, the CPU cost is about the same regardless of data or ACK. This means that the application which is making the propper gathering send call will spend far fewer CPU cycles in the networking stack. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, Oct 26, 2011 at 10:56 AM, Rick Jones <rick.jones2@hp.com> wrote: > On 10/25/2011 07:25 PM, Andy Lutomirski wrote: >> >> When talking to an unfixable interactive peer that fails to set >> TCP_NODELAY, disabling delayed ACKs can help mitigate the problem. >> This is an evil thing to do, but if the entire network is private, >> it's not that evil. >> >> This works around a problem with the remote *application*, so make >> it a socket option instead of a sysctl or a per-route option. >> >> Signed-off-by: Andy Lutomirski<luto@amacapital.net> >> --- >> >> This patch is a bit embarrassing. We talk to remote applications over >> TCP that are very much interactive but don't set TCP_NODELAY. These >> applications apparently cannot be fixed. As a partial workaround, if we >> ACK every incoming segment, then as long as they don't transmit two >> segments per rtt, we do pretty well. > > Embarrassing/evil indeed - is it really something to go into the kernel? That's a good question. It's in our kernel -- I don't know whether it should go upstream. > > If the networks where this happens are indeed truly private, can they run a > private kernel? Or use an LD_PRELOAD hack to wedge-in a > setsockopt(TCP_NODELAY) call into the application? Or set something like > tcp_naglim_def on the application system(s)? Or have the server application > make a setsockopt(TCP_MAXSEG) call before listen() to a value one byte below > that of what the application is sending? We control our server. We don't control the server at the other end. We've tried to get them to do any of the above, but they seem unwilling or unable to do it. I suspect that they're using various pieces from various third-party vendors that just don't care. > > Is the application actually "virtuous" in sending logically associated data > in one "send" call, and simply running afoul of Nagle+DelayedACK in having > multiple distinct requests outstanding at once, or is it actually quite evil > in that it is sending logically associated data in separate send calls? > The remote application generates messages meant for us, and they appear to send each message in its own segment. I don't have the source, so I don't know whether they're really using one send call per message or whether they're using MSG_MORE, TCP_CORK, so some other mechanism. Each message is time-sensitive and should be received as soon as possible afterq its sent (i.e. one-half rtt). Unfortunately, when they send two messages and we don't ack the first one, the second gets delayed. Turning off delayed acks helps but does not completely solve the problem. > rick jones > > choir preaching follows: :) I agree. Unfortunately I didn't write all this stuff. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
>> If the networks where this happens are indeed truly private, can they run a >> private kernel? Or use an LD_PRELOAD hack to wedge-in a >> setsockopt(TCP_NODELAY) call into the application? Or set something like >> tcp_naglim_def on the application system(s)? Or have the server application >> make a setsockopt(TCP_MAXSEG) call before listen() to a value one byte below >> that of what the application is sending? > > We control our server. We don't control the server at the other end. > We've tried to get them to do any of the above, but they seem > unwilling or unable to do it. I suspect that they're using various > pieces from various third-party vendors that just don't care. Making the setsockopt(TCP_MAXSEG) would be at your end :) Presumably based on the minimum message size. That would cause the connection to have an MSS == the request size so every request send should take the "is this send plus any queued unsent data >= MSS" path. Another "at your end" possibility would be setting a rather small SO_RCVBUF size at your end before calling listen(), in hopes of triggering the window update. >> Is the application actually "virtuous" in sending logically associated data >> in one "send" call, and simply running afoul of Nagle+DelayedACK in having >> multiple distinct requests outstanding at once, or is it actually quite evil >> in that it is sending logically associated data in separate send calls? >> > > The remote application generates messages meant for us, and they > appear to send each message in its own segment. I don't have the > source, so I don't know whether they're really using one send call per > message or whether they're using MSG_MORE, TCP_CORK, so some other > mechanism. Each message is time-sensitive and should be received as > soon as possible afterq its sent (i.e. one-half rtt). Unfortunately, > when they send two messages and we don't ack the first one, the second > gets delayed. Turning off delayed acks helps but does not completely > solve the problem. If it is write,write,read (multiple sends per logical message) in a packet trace you should see a partial request in the first segment, followed by the rest of the request (and perhaps the second through Nth) in the second segment. Or, I suppose your server application would have a receive complete with the first part of the first request, getting the second part of the request in a subsequent receive call. If it is multiple requests at a time each sent in one send call, you should see a first segment arriving with a complete request within it, followed by a second segment with the next request(s). rick jones -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, Oct 26, 2011 at 1:06 PM, Rick Jones <rick.jones2@hp.com> wrote: >>> If the networks where this happens are indeed truly private, can they run >>> a >>> private kernel? Or use an LD_PRELOAD hack to wedge-in a >>> setsockopt(TCP_NODELAY) call into the application? Or set something like >>> tcp_naglim_def on the application system(s)? Or have the server >>> application >>> make a setsockopt(TCP_MAXSEG) call before listen() to a value one byte >>> below >>> that of what the application is sending? >> >> We control our server. We don't control the server at the other end. >> We've tried to get them to do any of the above, but they seem >> unwilling or unable to do it. I suspect that they're using various >> pieces from various third-party vendors that just don't care. > > Making the setsockopt(TCP_MAXSEG) would be at your end :) Presumably based > on the minimum message size. That would cause the connection to have an MSS > == the request size so every request send should take the "is this send plus > any queued unsent data >= MSS" path. That's cute. The messages are variable-size (but they don't vary much), so doing this would probably be worse for the network than having them set TCP_NODELAY or having us turn off delayed acks, but we don't really care about the network, and it might work well. > > Another "at your end" possibility would be setting a rather small SO_RCVBUF > size at your end before calling listen(), in hopes of triggering the window > update. That scares me. If they every start sending in bursts (it happens on occasion), then we lose if they would want to exceed an artificially small window. > >>> Is the application actually "virtuous" in sending logically associated >>> data >>> in one "send" call, and simply running afoul of Nagle+DelayedACK in >>> having >>> multiple distinct requests outstanding at once, or is it actually quite >>> evil >>> in that it is sending logically associated data in separate send calls? >>> >> >> The remote application generates messages meant for us, and they >> appear to send each message in its own segment. I don't have the >> source, so I don't know whether they're really using one send call per >> message or whether they're using MSG_MORE, TCP_CORK, so some other >> mechanism. Each message is time-sensitive and should be received as >> soon as possible afterq its sent (i.e. one-half rtt). Unfortunately, >> when they send two messages and we don't ack the first one, the second >> gets delayed. Turning off delayed acks helps but does not completely >> solve the problem. > > If it is write,write,read (multiple sends per logical message) in a packet > trace you should see a partial request in the first segment, followed by the > rest of the request (and perhaps the second through Nth) in the second > segment. Or, I suppose your server application would have a receive > complete with the first part of the first request, getting the second part > of the request in a subsequent receive call. > > If it is multiple requests at a time each sent in one send call, you should > see a first segment arriving with a complete request within it, followed by > a second segment with the next request(s). These are asynchronous messages and we don't reply to the vast majority of them. We see one request arriving per segment. I'll play with TCP_MAXSEG. But I'll probably leave TCP_NO_DELAYED_ACK patched in to my kernel for the time being. I'm not thrilled about forcing the other side to split their messages across multiple segments. --Andy -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Le mardi 25 octobre 2011 à 19:25 -0700, Andy Lutomirski a écrit : > When talking to an unfixable interactive peer that fails to set > TCP_NODELAY, disabling delayed ACKs can help mitigate the problem. > This is an evil thing to do, but if the entire network is private, > it's not that evil. > > This works around a problem with the remote *application*, so make > it a socket option instead of a sysctl or a per-route option. > > Signed-off-by: Andy Lutomirski <luto@amacapital.net> > --- > > This patch is a bit embarrassing. We talk to remote applications over > TCP that are very much interactive but don't set TCP_NODELAY. These > applications apparently cannot be fixed. As a partial workaround, if we > ACK every incoming segment, then as long as they don't transmit two > segments per rtt, we do pretty well. > > Windows can do something similar, but it's per interface instead of per > socket: > > http://support.microsoft.com/kb/328890 Hi Andy Yet another delayed ack hacking proposal :) Well, to be honest, I find the MS Windows tunable more generic. [ But doing it for a whole interface is wrong, it should be per socket to allow best tuning ] Setting the value to 4 (instead of default 2) for example would _reduce_ number of ACK packets in bulk transferts [ We can do that if GRO is on, as a side effect ] Also the 40ms/200ms values (TCP_DELACK_{MIN|MAX}) could be tunables. (system or per socket) RFC 1122 says it SHOULD be less than 500ms. The time criteria is IMHO far more palatable for an application author than "number of delayed acks" -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Oct 27, 2011 at 1:24 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote: > Le mardi 25 octobre 2011 à 19:25 -0700, Andy Lutomirski a écrit : >> When talking to an unfixable interactive peer that fails to set >> TCP_NODELAY, disabling delayed ACKs can help mitigate the problem. >> This is an evil thing to do, but if the entire network is private, >> it's not that evil. >> >> This works around a problem with the remote *application*, so make >> it a socket option instead of a sysctl or a per-route option. >> >> Signed-off-by: Andy Lutomirski <luto@amacapital.net> >> --- >> >> This patch is a bit embarrassing. We talk to remote applications over >> TCP that are very much interactive but don't set TCP_NODELAY. These >> applications apparently cannot be fixed. As a partial workaround, if we >> ACK every incoming segment, then as long as they don't transmit two >> segments per rtt, we do pretty well. >> >> Windows can do something similar, but it's per interface instead of per >> socket: >> >> http://support.microsoft.com/kb/328890 > > Hi Andy > > Yet another delayed ack hacking proposal :) > > Well, to be honest, I find the MS Windows tunable more generic. > [ But doing it for a whole interface is wrong, it should be per socket > to allow best tuning ] > > Setting the value to 4 (instead of default 2) for example would _reduce_ > number of ACK packets in bulk transferts [ We can do that if GRO is on, > as a side effect ] > > Also the 40ms/200ms values (TCP_DELACK_{MIN|MAX}) could be tunables. > (system or per socket) > RFC 1122 says it SHOULD be less than 500ms. The time criteria is IMHO > far more palatable for an application author than "number of delayed > acks" Hello Eric, Few days ago, in our custom kernel we made TCP Delack segments and TCP Delack timeout parameters tunable via proc entries. Increasing tcp_delack_segs (number of full sized segments that must be received until an ACK is sent) we observed an improvement of throughput up to 20% in some test cases. Do you think that this kind of patch would have a chance to be included in mainstream? thanks, Daniel. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Le jeudi 27 octobre 2011 à 14:54 +0300, Daniel Baluta a écrit : > Few days ago, in our custom kernel we made TCP Delack segments and > TCP Delack timeout parameters tunable via proc entries. > Increasing tcp_delack_segs (number of full sized segments that must > be received until an ACK is sent) we observed an improvement of > throughput up to 20% in some test cases. > > Do you think that this kind of patch would have a chance to be > included in mainstream? > If your patches are ready, why not sending them as RFC ? -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Oct 27, 2011 at 3:13 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote: > Le jeudi 27 octobre 2011 à 14:54 +0300, Daniel Baluta a écrit : > >> Few days ago, in our custom kernel we made TCP Delack segments and >> TCP Delack timeout parameters tunable via proc entries. >> Increasing tcp_delack_segs (number of full sized segments that must >> be received until an ACK is sent) we observed an improvement of >> throughput up to 20% in some test cases. >> >> Do you think that this kind of patch would have a chance to be >> included in mainstream? >> > > If your patches are ready, why not sending them as RFC ? OK, I will port them to the latest kernel and send a patch. thanks, Daniel. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/include/linux/tcp.h b/include/linux/tcp.h index 531ede8..2116f31 100644 --- a/include/linux/tcp.h +++ b/include/linux/tcp.h @@ -106,6 +106,7 @@ enum { #define TCP_THIN_LINEAR_TIMEOUTS 16 /* Use linear timeouts for thin streams*/ #define TCP_THIN_DUPACK 17 /* Fast retrans. after 1 dupack */ #define TCP_USER_TIMEOUT 18 /* How long for loss retry before timeout */ +#define TCP_NO_DELAYED_ACK 19 /* Do not delay ACKs. */ /* for TCP_INFO socket option */ #define TCPI_OPT_TIMESTAMPS 1 diff --git a/include/net/inet_connection_sock.h b/include/net/inet_connection_sock.h index e6db62e..1ad91bf 100644 --- a/include/net/inet_connection_sock.h +++ b/include/net/inet_connection_sock.h @@ -106,8 +106,9 @@ struct inet_connection_sock { struct { __u8 pending; /* ACK is pending */ __u8 quick; /* Scheduled number of quick acks */ - __u8 pingpong; /* The session is interactive */ __u8 blocked; /* Delayed ACK was blocked by socket lock */ + __u8 pingpong:1; /* The session is interactive */ + __u8 nodelack:1; /* Delayed ACKs are disabled */ __u32 ato; /* Predicted tick of soft clock */ unsigned long timeout; /* Currently scheduled timeout */ __u32 lrcvtime; /* timestamp of last received data packet */ diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c index 46febca..e8e98dc 100644 --- a/net/ipv4/tcp.c +++ b/net/ipv4/tcp.c @@ -2385,6 +2385,13 @@ static int do_tcp_setsockopt(struct sock *sk, int level, } break; + case TCP_NO_DELAYED_ACK: + if (val == 0 || val == 1) + icsk->icsk_ack.nodelack = !!val; + else + err = -EINVAL; + break; + #ifdef CONFIG_TCP_MD5SIG case TCP_MD5SIG: /* Read the IP->Key mappings from userspace */ @@ -2564,6 +2571,10 @@ static int do_tcp_getsockopt(struct sock *sk, int level, val = !icsk->icsk_ack.pingpong; break; + case TCP_NO_DELAYED_ACK: + val = icsk->icsk_ack.nodelack; + break; + case TCP_CONGESTION: if (get_user(len, optlen)) return -EFAULT; diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index 21fab3e..e7d7ee0 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -197,7 +197,8 @@ static void tcp_enter_quickack_mode(struct sock *sk) static inline int tcp_in_quickack_mode(const struct sock *sk) { const struct inet_connection_sock *icsk = inet_csk(sk); - return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; + return (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong) || + icsk->icsk_ack.nodelack; } static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
When talking to an unfixable interactive peer that fails to set TCP_NODELAY, disabling delayed ACKs can help mitigate the problem. This is an evil thing to do, but if the entire network is private, it's not that evil. This works around a problem with the remote *application*, so make it a socket option instead of a sysctl or a per-route option. Signed-off-by: Andy Lutomirski <luto@amacapital.net> --- This patch is a bit embarrassing. We talk to remote applications over TCP that are very much interactive but don't set TCP_NODELAY. These applications apparently cannot be fixed. As a partial workaround, if we ACK every incoming segment, then as long as they don't transmit two segments per rtt, we do pretty well. Windows can do something similar, but it's per interface instead of per socket: http://support.microsoft.com/kb/328890 include/linux/tcp.h | 1 + include/net/inet_connection_sock.h | 3 ++- net/ipv4/tcp.c | 11 +++++++++++ net/ipv4/tcp_input.c | 3 ++- 4 files changed, 16 insertions(+), 2 deletions(-)