Imagine I have a server that can produce messages at a rate of 10,000 messages per second. But my client can only receive up to a maximum of 1000 messages per second.
System 1
If my system sends 1000 messages in the 1st milisecond and then does nothing for the remaining 999ms.
System 2
My system sends 1 message per milisecond, hence in 1000ms (1second) it will send 1000 messages.
Q1) Which system is better given that the client can handle a maximum of 500 messages per second?
Q2) What will be the impact of system 1 on the client? Will it overwhelm the client?
Thanks
Wil it overwhelm the client: It depends of the size of your messages, and the socket buffer size. The messages the sender sends are buffered. If the client cannot consume because the buffer is full, the output stream the sender is using will block. When the client has consumed some messages, the sender can continue writing as his OutputStream gets unblocked.
A typical buffer size on a windows system used to be 8192 bytes, but size can differ by the OS and settings in the OS.
So System 1 will not overwhelm the client, it will block at a certain moment.
What is the best approach merely depends on the design of your application.
For example: I had a similar issue while writing to an Arduino via USB (not socket-client, but otherwise the same problem). In my problem, buffered messages where a problem because it were positions of a face tracking camera. Buffered positions were no longer relevant when the Arduino read them, but it MUST process them because such a buffer is a queue, and you can only get the most recent if you read out the old one's. The Arduino could never keep up with the messages being produced, because by the time a new position reached the Arduino code, it was outdated. So that was an "overwhelm".
I resolved this by using bi-directional communication. The Arduino would send a message to the producer saying: READY (to receive a message). Then the producer would send one (up-to-date) face tracking position. Then the Arduino repositioned the camera and requested a new message. This way, there was a kind of flow control, that prevented the producer to overflow the Arduino.
Neither is better. TCP will alter the actual flow whatever you do yourself.
Neither will overwhelm the client. If the client isn't keeping up, its socket receive buffer will fill up, and so wil your socket send buffer, and eventually you will block in send, or get EAGAIN/EWOULDBLOCK if you're in non-blocking mode.
Related
I have a simple Java Server/Client TCP Socket program that calculates the rtt. The way I did it was I let the client time-stamp its outgoing message, then the server adds a time-stamp when it receives the message and finally when the message is returned to the client it adds another time-stamp.
Well, the result I ended up with is kind of weird. I made a plot of the message size vs round-trip time and ended up with a straight line rather than a linear increase.
Can someone explain why that happened?
First of all, I don't think you need to timestamp the message on the server side. You just need the sending time to be added when you send the packet and then you can calculate the difference with the current time when the response get's back to the client (of course the server needs to copy the sent_time in the reply).
note: If your server takes "long" time to execute then you will need to add some kind of "time to process" information to the packet by the server.
RTT = (cur_clientTime_when_response_is_received - time_sent) - time_to_server_process
This formula has the advantage that doesn't care about time differences between the client and the server clock.
The straight line is explainable because the difference in time to send the data is not statistically relevant to overcome the randomness introduced in the network layer as it's not a real time media, and because it's shared with other application and hosts, if you are on a network switch.
Anyway, don't bother increasing the number of letters, the underlaying protocol will take a max of MTU (that usually is 1500) bytes per TCP packet: still not enough.
We have a system which receives data from the users and pushes data to kafka and only when we are sure that the data has been pushed we send the user an "OK" response.
Since the new kafka is using async send(ProducerRecord,Callback), I wanted to know that if this send is crash resistant (fault-tolerant)?
My guess is that its most probably not,so how can I use it in sync mode? Or should I make the user wait until the callback is called?
According to Kafka's Design :
Asynchronous send
Batching is one of the big drivers of efficiency, and to enable batching the Kafka producer has an asynchronous mode that accumulates data in memory and sends out larger batches in a single request. The batching can be configured to accumulate no more than a fixed number of messages and to wait no longer than some fixed latency bound (say 100 messages or 5 seconds). This allows the accumulation of more bytes to send, and few larger I/O operations on the servers. Since this buffering happens in the client it obviously reduces the durability as any data buffered in memory and not yet sent will be lost in the event of a producer crash.
I have a Java program that is constantly being sent UDP data from an external system.
Periodically, we need to stop receiving data (because another machine is handling it). During those times, my socket reader thread goes into a sleep loop. When it is time to start receiving packets, I go into socket.receive(Packet) again and have a buffer full of packets that I should not be handling. (The data came while in the "stop time".)
Is there a way to clear the buffer of a DatagramSocket?
If not, what is the best alternative? Set the buffer size to 0 when I go into the wait state and bring it back when I start to service packets again? Close the socket when I while I wait and open a new one when I come back?
Rather than having the downtime on the socket, have the downtime on whatever code processes the packets.
So the socket continues to receive like it normally would, but if it's on downtime, it just immediately drops the packet.
Not exactly the most efficient solution, but it's really easy to implement and might be useful as it leaves the node open in other cases for accepting different types of packets at different times.
I develop part of some JBoss+EJB based enterprise application. My module needs to process huge amount of incoming UDP packets. I've done some load testing and it looks that in case of sending packets with 11ms interval everything is fine, but in case of 10ms interval some packets are lost. It's rather strange in my opinion, but I done 10/11ms interval load tests comparison several times and it is always the same result (10 ms - some "lost" packets, 11ms - everything's fine).
If it was something wrong with synchronization, I'd expect that it will also be visible in case of 11ms tests (at least one packet lost, or at least one wrong counter value).
So if it is not because of synchronization, then maybe DatagramSocket through which I receive packets doesn't work as expected.
I found that receive buffer size (SO_RCVBUF) has default 57344 value (probably it's underlying IO network buffers dependent). I suspect, that maybe when this buffer goes full, then new incoming UDP datagrams are rejected. I tried set this value to some higher, but I noticed that if I exaggerate, buffer returns to its default size. If it's underlying layer dependent how can I find out maximum buffer size for certain OS/network card from JBoss level?
Is it possible that it is caused by receive buffer size, or maybe 57344 value is big enough to handle most cases? Do you have any experience with such issues?
There is no timeout set on my DatagramSocket. My UDP datagrams contains about 70 bytes of data (value without datagram header included).
[Edited]
I have to use UDP because I receive Cisco Netflow data - it is protocol used by network devices to send some traffic statistics. Also, I have no influence on sent bytes format (e.g. I cannot add counters for packets and so on). It is not expected that all packets will be processed (some datagrams may be lost), but I'd expect that I will process most of packets. During 10ms interval tests, about 30% of packets were lost.
It is not very possible that slow processing causes this issue. Currently singleton component holds reference to DatagramSocket calling receive method in a loop. When packet is received, it is passed to the queue, and processed by picked from pool stateless component. "Facade" Singleton is only responsible for receiving packets and passing it on to the processing (it does not wait for processing complete event).
Thanks in advance,
Piotr
UDP does not guarantee delivery, so you can tweak parameters, but you can't guarantee that the message will get delivered, especially in the case of very large data transfers.
If you need to guarantee delivery, you should use TCP instead.
If you need (or want) to use UDP, you can encode each packet with a number, and also send the number of packets expected. For example, if you sent 10 large packets, you could include the information: packet 1/10, packet 2/10, etc. This way you can at least tell if you have not received all of the packets. If you have not received them, you could send a request to resend those missing packets.
UDP is inherently unreliable.
Datagrams can be thrown away at any point between sender and receiver, even within the receiver at a level below your code. Setting the recv buffer to a larger size is likely to help the networking code within your machine buffer more datagrams but you should expect that some datagrams will be lost anyway.
If your recv logic takes too long (i.e. longer than it takes for a new datagram to arrive) then you'll always be behind and you'll always miss datagrams eventually. All you can do is make sure that your recv code runs as fast as possible, perhaps move the inbound datagram to a queue and process it 'later' or on another thread but then that will just move your problem to being one where you have a queue that keeps growing.
[Re your edit...] And what's processing your queue and how does the locking work between the producer and the consumers? Change your code so that the recv logic simply increments a count and discards the data and loops back around and see if you're losing fewer datagrams; either way, UDP is unreliable, you WILL have datagrams that are discarded and you should just expect that and deal with it. Worrying about it means you're focusing on the wrong problem; make use of the data you DO get and assume that you wont get much of it and then your program will work even if the network gets congested and MOST of your datagrams get discarded.
In summary, that's just how is it with UDP.
It appears in your tests that only up to two packets can be in the buffer so if each packet is less than 28KB this should be fine.
As you know UDP is lossy, but you should be able to send more than one packet per 10 ms. I suggest you write a simple receiver which just listens to packets just to determine if its your application or something at the network/OS level. (I suspect the later)
I don't know Java but ... does the API allow you to invoke an asynch listen/receive for a datagram:
Use O/S API to do a receive (passing your application-level buffer as a paremeter)
(Wait while there's nothing to receive...)
(O/S receives something from the network...)
O/S puts the received packet into the buffer and completes/returns your API call
If that's true then I suggest that you do several concurrent instances of the API call, so that there are several concurrent application-level buffers into which multiple packets can be received.
I'm making my own custom server software for a game in Java (the game and original server software were written with Java). There isn't any protocol documentation available, so I am having to read the packets with Wireshark.
While a client is connecting the server sends it the level file in Gzip format. At about 94 packets into sending the level, my server crashes the client with an ArrayIndexOutOfBoundsException. According to the capture file from the original server, it sends a TCP Window Update at about that point. What is a TCP Window Update, and how would I send one using a SocketChannel?
TCP windows are used for flow control between the peers on a connection. With each ACK packet, a host will send a "window size" field. This field says how many bytes of data that host can receive before it's full. The sender is not supposed to send more than that amount of data.
The window might get full if the client isn't receiving data fast enough. In other words, the TCP buffers can fill up while the application is off doing something other than reading from its socket. When that happens, the client would send an ACK packet with the "window full" bit set. At that point, the server is supposed to stop sending data. Any packets sent to a machine with a full window will not be acknowledged. (This will cause a badly behaved sender to retransmit. A well-behaved sender will just buffer the outgoing data. If the buffer on the sending side fills up too, then the sending app will block when it tries to write more data to the socket!)
This is a TCP stall. It can happen for a lot of reasons, but ultimately it just means the sender is transmitting faster than the receiver is reading.
Once the app on the receiving end gets back around to reading from the socket, it will drain some of the buffered data, which frees up some space. The receiver will then send a "window update" packet to tell the sender how much data it can transmit. The sender starts transmitting its buffered data and traffic should flow normally.
Of course, you can get repeated stalls if the receiver is consistently slow.
I've worded this as if the sender and receiver are different, but in reality, both peers are exchanging window updates with every ACK packet, and either side can have its window fill up.
The overall message is that you don't need to send window update packets directly. It would actually be a bad idea to spoof one up.
Regarding the exception you're seeing... it's not likely to be either caused or prevented by the window update packet. However, if the client is not reading fast enough, you might be losing data. In your server, you should check the return value from your Socket.write() calls. It could be less than the number of bytes you're trying to write. This happens if the sender's transmit buffer gets full, which can happen during a TCP stall. You might be losing bytes.
For example, if you're trying to write 8192 bytes with each call to write, but one of the calls returns 5691, then you need to send the remaining 2501 bytes on the next call. Otherwise, the client won't see the remainder of that 8K block and your file will be shorter on the client side than on the server side.
This happens really deep in the TCP/IP stack; in your application (server and client) you don't have to worry about TCP windows. The error must be something else.
TCP WindowUpdate - This indicates that the segment was a pure WindowUpdate segment. A WindowUpdate occurs when the application on the receiving side has consumed already received data from the RX buffer causing the TCP layer to send a WindowUpdate to the other side to indicate that there is now more space available in the buffer. Typically seen after a TCP ZeroWindow condition has occurred. Once the application on the receiver retrieves data from the TCP buffer, thereby freeing up space, the receiver should notify the sender that the TCP ZeroWindow condition no longer exists by sending a TCP WindowUpdate that advertises the current window size.
https://wiki.wireshark.org/TCP_Analyze_Sequence_Numbers
A TCP Window Update has to do with communicating the available buffer size between the sender and the receiver. An ArrayIndexOutOfBoundsException is not the likely cause of this. Most likely is that the code is expecting some kind of data that it is not getting (quite possibly well before this point that it is only now referencing). Without seeing the code and the stack trace, it is really hard to say anything more.
You can dive into this web site http://www.tcpipguide.com/free/index.htm for lots of information on TCP/IP.
Do you get any details with the exception?
It is not likely related to the TCP Window Update packet
(have you seen it repeat exactly for multiple instances?)
More likely related to your processing code that works on the received data.
This is normally just a trigger, not the cause of your problem.
For example, if you use NIO selector, a window update may trigger the wake up of a writing channel. That in turn triggers the faulty logic in your code.
Get a stacktrace and it will show you the root cause.