The strange thing is that even there are a lot physical memory available.

That is the reason I think the virtual memory was used by the socket. As you mentioned it only consume small amount of physical memory. all of this should be all assigned to physical memory. When my program start new thread, it get a new a tcp socket from connection( in C++ it is a integer to represent the connection) and 4 4KB buffer for socket read/write and encrypt/decrypt plus 32 byte AES key variable. Which is fine, until everyone tries to use it at the same time. When in reality there is not that much memory physically available. That means that many processes/threads could all ask for a lot of memory and all be told 'OK' they have it.

BUT, that memory space may not be mapped from real physical RAM into the applications virtual address space until it is actually used by the application later. The kernel will say 'yep OK, here it is'. To make things more interesting, when you application starts a new thread it may well ask the kernel for 8MB or whatever of memory for its stack space and so on. There is a mapping going on between virtual and physical addresses. The memory addresses your application sees are not actual physical memory addresses. Use HoldForm to leave the expression unevaluated: ( TraditionalForm uses classic mathematical typesetting. A setting of 1 approaches the limit from the right: In 2. A setting of 1 approaches the limit from the left: In 1. The thing is Linux employs 'virtual memory'. You can also specify the limit’s Direction. It seems to me the OS Socket operation is based on the file system IO

At least, not obviously unreasonable.It is also strange to me, when the program started new thread(connection), the virtual memory usage get increased not the real physical memory. Not as fast as the tree code, but still (I think) reasonable behavior. Here is Leonid Shifrin's example: Timing]]]] The following will build a binary tree, and then analyze it and extract the results: Clear

Wolfram mathematica limit memory