[yosefe]

Why

The UCT sockaddr negative-connect test connects to a non-existing IP address and waits for the client error callback. On hpc-build-ub, the TCP CM variants were taking about 135 seconds each because the socket remained in SYN retransmission timeout handling. The 9 TCP parameters in tcp/test_uct_sockaddr_err_handle_non_exist_ip.conn_to_non_exist_ip consumed about 20 minutes while still covering only the same negative-connect path.

The test attempted to shorten the wait with SYN_CNT, but the TCP CM configuration key needs the TCP_CM_ prefix when modifying the config bundle, so the setting was not applied to the CM socket.

How

Add a TCP USER_TIMEOUT config option for both TCP iface and TCP sockcm sockets. When set, UCX applies Linux TCP_USER_TIMEOUT on a best-effort basis, guarded by #ifdef TCP_USER_TIMEOUT so non-Linux builds continue to compile and keep the default behavior.

The negative-connect test now sets TCP_CM_SYN_CNT=1 and TCP_CM_USER_TIMEOUT=1s for transports that support those TCP CM options. This preserves the test coverage: it still connects to the non-existing IP and verifies the server-unavailable callback, but the kernel is asked to bound the failed connect attempt instead of waiting for the host default retry window.

Testing

On hpc-build-ub:

• module load dev/cuda-latest dev/gdrcopy-latest tools/hpc
• make -C src/uct -j16 libuct.la
• make -C test/gtest -j16 gtest
• ./test/gtest/gtest --gtest_filter="tcp/test_uct_sockaddr_err_handle_non_exist_ip.conn_to_non_exist_ip*"

Result: all 9 tests passed in 9258 ms. Before the fix, the same TCP cases took about 135 seconds each on this host.