Author: Wei Dong [email protected]

This is a stress test client of https://github.com/weiqj/pc12306.

The code depends on a C++11 compiler and the Boost library.

The program pc12306c-mpi.cpp is almost the same as pc12306c.cpp except for MPI support. A MPI_Barrier is used so that all clients start firing at the server at the same time. Use "make mpi" to compile.

  -h [ --help ]                    produce help message.
  -s [ --server ] arg (=localhost)
  -p [ --port ] arg (=12306)
  --trains arg (=5000)
  --segments arg (=10)
  --seats arg (=3000)
  --max-length arg (=5)
  -N arg (=10000000)               queries per client
  -C arg (=4)                      number parallel clients
  -B arg (=20)                     request batch size
  -Q arg (=1000)                   maximal outstanding request per client
  -S arg (=100)                    if queue is fall, sleep this # us
  --cycle arg (=1)                 print counters every cycle seconds

Explanation of some parameters:

• C: each client runs two threads, one for reading and one for writing, so total thread number is 2C.
• B: fit multiple requests into a single send request might increase throughput. (With 1GB ethernet, it seems B doesn't help much in improving throughput.)
• Q: this is to preventing sending out too many requests before they can be processed by server. If outstanding query + B >= Q, the writing thread will sleep S us and try again. Larger Q helps improve throughput.
• N: this program pre-allocate memory for all requests and responses. For N=10000000 and C=4 (default), the maximal resident set size of the program is about 4GB. The memory consumption is proportional to NxC.

• ooo.time.max: ooo == out of order. Usually smaller respID should arrive before bigger respID. This value is the maximal time a bigger respID has arrived before a smaller respID.
• fill.rate: the server keeps the state across multiple invocation of the client. This value is meaningful only on the first invocation of client after server restart, because subsequent invocations of client have no means to track how many seats were sold before. This value of the first N invocation of client can be added up for the total rate of seats sold.

There's some Heisenbug causing the program to crash when evaluation conflict1 and/or conflict2.