Note that you are sending integers in host byte order, which means you are subject to endian differences between the client and server. You really should be using network byte order instead. If you are not going to use the IOHandler's Read/Write methods for integer values, at least use Indy's GStack.HostToNetwork() function before sending the data, and the GStack.NetworkToHost() function after receiving the data, to make sure the integers are in a consistent endian on the wire.
PUInt32(@ sendBuf [0])^ := GStack.HostToNetwork(UInt32(CMDCOMMU_HEADER)); PUInt32(@ sendBuf [4])^ := GStack.HostToNetwork(UInt32(bufsize)); ... PUInt32(@ arrBuf [0])^ := GStack.NetworkToHost(PUInt32(@ arrBuf [0])^); PUInt32(@ arrBuf [4])^ := GStack.NetworkToHost(PUInt32(@ arrBuf [4])^);
Otherwise, use the IOHandler's reading/writing functions: with cmdClient.fIdClient.IOHandler do begin Write(Int32(CMDCOMMU_HEADER)); Write(Int32(bufsize)); Write(buf); end; ... with argContext.Connection.IOHandler do begin PktType := ReadInt32; PktSize := ReadInt32; ReadBytes(arrBuf, PktSize); end;
Do uses GStack.HostToNetwork to prevent some network device reading error?
What is the point of putting a fixed header on the packets if your server is not going to use the header for reading the packet? You really shouldn't be doing arbitrary reading in this situation. You know the exact data size up front, so just ask Indy to read that many bytes. Problem solved.
TCP is a byte stream, it doesn't matter if the data is tranmitted on the wire in 1 packet or 100 packets. TCP guarantees reliable delivery. The bytes will be accurate on the receiving end, no matter how many packets it has to read. But, if you want to reduce the number of TCP packets transmitted, you can use Indy's write buffering when sending a packet, eg:
with cmdClient.fIdClient.IOHandler do begin WriteBufferOpen; try Write(...); Write(...); Write(...); WriteBufferClose; except WriteBufferCancel; raise; end end;
Freq, client sends two times, server gets just one time.
Some real time requirement
TIdTCPClient and TIdTCPServer use TCP, hence their names. Do you understand how TCP actually works? Do you understand that there is no 1-to-1 relationship between sends and reads in TCP (unlike in UDP)? You could sent 100 bytes 1 time, but read 1 byte 100 times. TCP guarantees the bytes are delivered, and in the right order, but how you send the bytes does not affect how you read the bytes and vice versa.
In any case, regarding the 2-sends 1-read issue, it likely has to do with how ReadBytes(-1) interacts with the InputBuffer. After the OnExecute event exits, the server calls IOHandler.Connected before firing OnExecute again. Connected performs a read, so it is possible that bytes are getting read into the InputBuffer and a subsequent ReadBytes(-1) is not checking for them before waiting for more bytes to arrive on the socket. If you were reading the packet header, like you should be, you wouldn't have that issue.
It is just like you told me before @rlebeau
Toggles between ReadBytes(8) and ReadBytes(argConn.bufSize)
TIdTCPClient.Connected is not thread safe, I think.
A mutex needs to be added on TIdTCPClient.IoHandler.ReadBytes call in the receive thread and on TIdTCPClient.Connected in the connection thread.
I found that the client receive thread can be aborted by the exception of the calling IoHandler.ReadBytes, then the problem solved.
A mutex needs to be added on TIdTCPClient.IoHandler.ReadBytes call in the receive thread and on TIdTCPClient.Connected in the connection thread.
I found that the client receive thread can be aborted by the exception of the calling IoHandler.ReadBytes, then the problem solved.
If you don't want the thread to terminate, catch the exception
What Is wrong in this implementation?
In the client, I start a receive thread:
--the receive thread in the client:
----TIdTCPClient.IoHandler.ReadBytes
in the server:
--OnExcute:
----AContext.Connection.IoHandler.ReadBytes
If I first start the test transferring of the client sending data of < 80000 bytes to the server 10000 times and then do the same test reverse of the server sending to the client, there is nothing abnormal.
But if I first start the test of the server sending to the client and then the transferring speed becomes very low when do the test of the client sending to the server.
The test is one app sends data in size of < 7kb 100000 times to the other one.
If I never click the Server Send button to start the test of the server sending to the client and just do test the client to the server, it takes about 40~50 seconds in every test.
But if I click the Server Send once, the test of the client to the server sending will take apparently longer time than before.
If I never click the Server Send button to start the test of the server sending to the client and just do test the client to the server, it takes about 40~50 seconds in every test.
But if I click the Server Send once, the test of the client to the server sending will take apparently longer time than before.
Here is the common read write procedure
The client
well, it kind of makes sense, if you take into account that the client code is single-threaded but the server code is multi-threaded, so you are likely getting task switches during the server-to-client sending that are slowing it down compared to the client-to-server sending. Remember, TIdTCPServer runs each client in its own worker thread. Even if there is only 1 client connected, there are still other threads running (the main UI thread, the server's listening thread, etc) that the OS has to service in a timely manner.
So I guess the non-fixed bytes sending causes some threads work so hard in TIdTCPServer if the TIdContext.Connection.IoHandler cached something when Write().
