Try the following to see if the (probably small) reduction in memory allocation / deallocation helps (I really doubt it will help much - I think there's something else that is causing problems, since you're seeing some major performance problems that probably have nothing to do with the code sample you posted): Change: IProtocolPacket* packet = iter->Next(); std::vector<asio::const_buffer> buffs(2); buffs.push_back(asio::const_buffer(packet->GetHeader(), packet->GetHeaderSize()); buffs.push_back(asio::const_buffer(packet->GetBody(), packet->GetBodySize()); To: IProtocolPacket* packet = iter->Next(); boost::array<asio::const_buffer, 2> buffs = { asio::buffer(packet->GetHeader(), packet->GetHeaderSize()), asio::buffer(packet->GetBody(), packet->GetBodySize()) }; (Warning - uncompiled and untested.) --- Hmmm - I see something that could be suspicious: "packet" is a local variable (short lifetime) - are the pointers obtained from packet valid (point to good memory) until the handle_send function is called? Asio does not own the data that the buffer objects wrap. Otherwise, you might want to post your io_service run code to see if there's something suspicious there. Cliff