diff options
| -rw-r--r-- | host/docs/transport.rst | 42 | ||||
| -rw-r--r-- | host/lib/transport/CMakeLists.txt | 10 | ||||
| -rw-r--r-- | host/lib/transport/udp_wsa_zero_copy.cpp | 281 |
3 files changed, 320 insertions, 13 deletions
diff --git a/host/docs/transport.rst b/host/docs/transport.rst index f28d1efcb..b601cd8ff 100644 --- a/host/docs/transport.rst +++ b/host/docs/transport.rst @@ -19,7 +19,10 @@ The transport parameters are defined below for the various transports in the UHD ------------------------------------------------------------------------ UDP transport (sockets) ------------------------------------------------------------------------ -The UDP transport is implemented with standard user-space/Berkeley sockets. +The UDP transport is implemented with user-space sockets: + +* **UNIX:** standard Berkeley sockets API using send()/recv() +* **Windows:** Windows Sockets API (WSA) using overlapped IO ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Transport parameters @@ -31,21 +34,18 @@ The following parameters can be used to alter the transport's default behavior: * **send_frame_size:** The size of a single send buffer in bytes * **num_send_frames:** The number of send buffers to allocate -**Note1:** num_recv_frames and num_send_frames do not affect performance. +**Note1:** +num_recv_frames does not affect performance (all platforms). + +**Note2:** +num_send_frames does not affect performance (UNIX only). -**Note2:** recv_frame_size and send_frame_size can be used to +**Note3:** +recv_frame_size and send_frame_size can be used to increase or decrease the maximum number of samples per packet. The frame sizes default to an MTU of 1472 bytes per IP/UDP packet, and may be increased if permitted by your network hardware. -**Note3:** For lower latency at low sample rates, use smaller buffers - -**Note4:** For overall lower latency, look for Interrupt Coalescing settings -for your OS and ethernet chipset. It seems the Intel ethernet chipsets offer -fine-grained control in Linux. Also, consult: - -* http://publib.boulder.ibm.com/infocenter/pseries/v5r3/index.jsp?topic=/com.ibm.aix.prftungd/doc/prftungd/interrupt_coal.htm - ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Flow control parameters ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ @@ -73,6 +73,25 @@ The following parameters can be used to alter socket's buffer sizes: **Note:** Large send buffers tend to decrease transmit performance. ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +Latency Optimization +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ +Latency is a measurement of the time it takes a sample to travel between the host and device. +Most computer hardware and software is bandwidth optimized which may negatively affect latency. +If your application has strict latency requirements, please consider the following notes: + +**Note1:** +The time taken by the device to populate a packet is proportional to the sample rate. +Therefore, to improve receive latency, configure the transport for a smaller frame size. + +**Note2:** +For overall latency improvements, +look for "Interrupt Coalescing" settings for your OS and ethernet chipset. +It seems the Intel ethernet chipsets offer fine-grained control in Linux. +Also, consult: + +* http://publib.boulder.ibm.com/infocenter/pseries/v5r3/index.jsp?topic=/com.ibm.aix.prftungd/doc/prftungd/interrupt_coal.htm + +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Linux specific notes ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ On linux, the maximum buffer sizes are capped by the sysctl values @@ -90,7 +109,6 @@ Windows specific notes ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ On Windows, it is important to change the default UDP behavior such that 1500 byte packets still travel through the fast path of the sockets stack. - FastSendDatagramThreshold registry key to change documented here: * http://www.microsoft.com/windows/windowsmedia/howto/articles/optimize_web.aspx#appendix_e diff --git a/host/lib/transport/CMakeLists.txt b/host/lib/transport/CMakeLists.txt index b1821956c..866ade75f 100644 --- a/host/lib/transport/CMakeLists.txt +++ b/host/lib/transport/CMakeLists.txt @@ -79,6 +79,15 @@ SET_SOURCE_FILES_PROPERTIES( PROPERTIES COMPILE_DEFINITIONS "${IF_ADDRS_DEFS}" ) +######################################################################## +# Setup UDP +######################################################################## +IF(WIN32) + LIBUHD_APPEND_SOURCES(${CMAKE_CURRENT_SOURCE_DIR}/udp_wsa_zero_copy.cpp) +ELSE() + LIBUHD_APPEND_SOURCES(${CMAKE_CURRENT_SOURCE_DIR}/udp_zero_copy.cpp) +ENDIF() + #On windows, the boost asio implementation uses the winsock2 library. #Note: we exclude the .lib extension for cygwin and mingw platforms. IF(WIN32) @@ -97,6 +106,5 @@ LIBUHD_APPEND_SOURCES( ${CMAKE_CURRENT_SOURCE_DIR}/buffer_pool.cpp ${CMAKE_CURRENT_SOURCE_DIR}/if_addrs.cpp ${CMAKE_CURRENT_SOURCE_DIR}/udp_simple.cpp - ${CMAKE_CURRENT_SOURCE_DIR}/udp_zero_copy.cpp ${CMAKE_CURRENT_SOURCE_DIR}/usb_zero_copy_wrapper.cpp ) diff --git a/host/lib/transport/udp_wsa_zero_copy.cpp b/host/lib/transport/udp_wsa_zero_copy.cpp new file mode 100644 index 000000000..ccfed38ea --- /dev/null +++ b/host/lib/transport/udp_wsa_zero_copy.cpp @@ -0,0 +1,281 @@ +// +// Copyright 2010-2011 Ettus Research LLC +// +// This program is free software: you can redistribute it and/or modify +// it under the terms of the GNU General Public License as published by +// the Free Software Foundation, either version 3 of the License, or +// (at your option) any later version. +// +// This program is distributed in the hope that it will be useful, +// but WITHOUT ANY WARRANTY; without even the implied warranty of +// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the +// GNU General Public License for more details. +// +// You should have received a copy of the GNU General Public License +// along with this program. If not, see <http://www.gnu.org/licenses/>. +// + +#include "udp_common.hpp" +#include <uhd/transport/udp_zero_copy.hpp> +#include <uhd/transport/udp_simple.hpp> //mtu +#include <uhd/transport/bounded_buffer.hpp> +#include <uhd/transport/buffer_pool.hpp> +#include <uhd/utils/msg.hpp> +#include <uhd/utils/log.hpp> +#include <boost/format.hpp> +#include <vector> + +using namespace uhd; +using namespace uhd::transport; +namespace asio = boost::asio; + +//A reasonable number of frames for send/recv and async/sync +static const size_t DEFAULT_NUM_FRAMES = 32; + +/*********************************************************************** + * Static initialization to take care of WSA init and cleanup + **********************************************************************/ +struct uhd_wsa_control{ + uhd_wsa_control(void){ + WSADATA wsaData; + WSAStartup(MAKEWORD(2, 2), &wsaData); /*windows socket startup */ + } + + ~uhd_wsa_control(void){ + WSACleanup(); + } +}; + +/*********************************************************************** + * Reusable managed receiver buffer: + * - Initialize with memory and a release callback. + * - Call get new with a length in bytes to re-use. + **********************************************************************/ +class udp_zero_copy_asio_mrb : public managed_recv_buffer{ +public: + udp_zero_copy_asio_mrb(void *mem, bounded_buffer<udp_zero_copy_asio_mrb *> &pending): + _mem(mem), _len(0), _pending(pending){/* NOP */} + + void release(void){ + if (_len == 0) return; + _pending.push_with_haste(this); + _len = 0; + } + + sptr get_new(size_t len){ + _len = len; + return make_managed_buffer(this); + } + + template <class T> T cast(void) const{return static_cast<T>(_mem);} + +private: + const void *get_buff(void) const{return _mem;} + size_t get_size(void) const{return _len;} + + void *_mem; + size_t _len; + bounded_buffer<udp_zero_copy_asio_mrb *> &_pending; +}; + +/*********************************************************************** + * Reusable managed send buffer: + * - committing the buffer calls the asynchronous socket send + * - getting a new buffer performs the blocking wait for completion + **********************************************************************/ +class udp_zero_copy_asio_msb : public managed_send_buffer{ +public: + udp_zero_copy_asio_msb(void *mem, int sock_fd, const size_t frame_size): + _sock_fd(sock_fd), _frame_size(frame_size), _committed(false) + { + _wsa_buff.buf = reinterpret_cast<char *>(mem); + ZeroMemory(&_overlapped, sizeof(_overlapped)); + _overlapped.hEvent = WSACreateEvent(); + UHD_ASSERT_THROW(_overlapped.hEvent != WSA_INVALID_EVENT); + this->commit(0); //makes buffer available via get_new + } + + ~udp_zero_copy_asio_msb(void){ + WSACloseEvent(_overlapped.hEvent); + } + + UHD_INLINE void commit(size_t len){ + if (_committed) return; + _committed = true; + _wsa_buff.len = len; + if (len == 0) WSASetEvent(_overlapped.hEvent); + else WSASend(_sock_fd, &_wsa_buff, 1, NULL, 0, &_overlapped, NULL); + } + + UHD_INLINE sptr get_new(const double timeout, size_t &index){ + const DWORD result = WSAWaitForMultipleEvents( + 1, &_overlapped.hEvent, true, DWORD(timeout*1000), true + ); + if (result == WSA_WAIT_TIMEOUT) return managed_send_buffer::sptr(); + index++; //advances the caller's buffer + + WSAResetEvent(_overlapped.hEvent); + _committed = false; + _wsa_buff.len = _frame_size; + return make_managed_buffer(this); + } + +private: + void *get_buff(void) const{return _wsa_buff.buf;} + size_t get_size(void) const{return _wsa_buff.len;} + + int _sock_fd; + const size_t _frame_size; + bool _committed; + WSAOVERLAPPED _overlapped; + WSABUF _wsa_buff; +}; + +/*********************************************************************** + * Zero Copy UDP implementation with WSA: + * + * This is not a true zero copy implementation as each + * send and recv requires a copy operation to/from userspace. + * + * For receive, use a blocking recv() call on the socket. + * This has better performance than the overlapped IO. + * For send, use overlapped IO to submit async sends. + **********************************************************************/ +class udp_zero_copy_wsa_impl : public udp_zero_copy{ +public: + typedef boost::shared_ptr<udp_zero_copy_wsa_impl> sptr; + + udp_zero_copy_wsa_impl( + const std::string &addr, + const std::string &port, + const device_addr_t &hints + ): + _recv_frame_size(size_t(hints.cast<double>("recv_frame_size", udp_simple::mtu))), + _num_recv_frames(size_t(hints.cast<double>("num_recv_frames", DEFAULT_NUM_FRAMES))), + _send_frame_size(size_t(hints.cast<double>("send_frame_size", udp_simple::mtu))), + _num_send_frames(size_t(hints.cast<double>("num_send_frames", DEFAULT_NUM_FRAMES))), + _recv_buffer_pool(buffer_pool::make(_num_recv_frames, _recv_frame_size)), + _send_buffer_pool(buffer_pool::make(_num_send_frames, _send_frame_size)), + _pending_recv_buffs(_num_recv_frames), + _next_send_buff_index(0) + { + UHD_MSG(status) << boost::format("Creating WSA UDP transport for %s:%s") % addr % port << std::endl; + static uhd_wsa_control uhd_wsa; //makes wsa start happen via lazy initialization + + UHD_ASSERT_THROW(_num_send_frames <= WSA_MAXIMUM_WAIT_EVENTS); + + //resolve the address + asio::io_service io_service; + asio::ip::udp::resolver resolver(io_service); + asio::ip::udp::resolver::query query(asio::ip::udp::v4(), addr, port); + asio::ip::udp::endpoint receiver_endpoint = *resolver.resolve(query); + + //create the socket + _sock_fd = WSASocket(AF_INET, SOCK_DGRAM, IPPROTO_UDP, NULL, 0, WSA_FLAG_OVERLAPPED); + if (_sock_fd == INVALID_SOCKET){ + const DWORD error = WSAGetLastError(); + throw uhd::os_error(str(boost::format("WSASocket() failed with error %d") % error)); + } + + //set the socket non-blocking for recv + u_long mode = 1; + ioctlsocket(_sock_fd, FIONBIO, &mode); + + //resize the socket buffers + const int recv_buff_size = int(hints.cast<double>("recv_buff_size", 0.0)); + const int send_buff_size = int(hints.cast<double>("send_buff_size", 0.0)); + if (recv_buff_size > 0) setsockopt(_sock_fd, SOL_SOCKET, SO_RCVBUF, (const char *)&recv_buff_size, sizeof(recv_buff_size)); + if (send_buff_size > 0) setsockopt(_sock_fd, SOL_SOCKET, SO_SNDBUF, (const char *)&send_buff_size, sizeof(send_buff_size)); + + //connect the socket so we can send/recv + const asio::ip::udp::endpoint::data_type &servaddr = *receiver_endpoint.data(); + if (WSAConnect(_sock_fd, (const struct sockaddr *)&servaddr, sizeof(servaddr), NULL, NULL, NULL, NULL) != 0){ + const DWORD error = WSAGetLastError(); + closesocket(_sock_fd); + throw uhd::os_error(str(boost::format("WSAConnect() failed with error %d") % error)); + } + + //allocate re-usable managed receive buffers + for (size_t i = 0; i < get_num_recv_frames(); i++){ + _mrb_pool.push_back(boost::shared_ptr<udp_zero_copy_asio_mrb>( + new udp_zero_copy_asio_mrb(_recv_buffer_pool->at(i), _pending_recv_buffs) + )); + _pending_recv_buffs.push_with_haste(_mrb_pool.back().get()); + } + + //allocate re-usable managed send buffers + for (size_t i = 0; i < get_num_send_frames(); i++){ + _msb_pool.push_back(boost::shared_ptr<udp_zero_copy_asio_msb>( + new udp_zero_copy_asio_msb(_send_buffer_pool->at(i), _sock_fd, get_send_frame_size()) + )); + } + } + + ~udp_zero_copy_wsa_impl(void){ + closesocket(_sock_fd); + } + + /******************************************************************* + * Receive implementation: + * + * Perform a non-blocking receive for performance, + * and then fall back to a blocking receive with timeout. + * Return the managed receive buffer with the new length. + * When the caller is finished with the managed buffer, + * the managed receive buffer is released back into the queue. + ******************************************************************/ + managed_recv_buffer::sptr get_recv_buff(double timeout){ + udp_zero_copy_asio_mrb *mrb = NULL; + if (_pending_recv_buffs.pop_with_timed_wait(mrb, timeout)){ + + ssize_t ret = ::recv(_sock_fd, mrb->cast<char *>(), _recv_frame_size, 0); + if (ret > 0) return mrb->get_new(ret); + + if (wait_for_recv_ready(_sock_fd, timeout)) return mrb->get_new( + ::recv(_sock_fd, mrb->cast<char *>(), _recv_frame_size, 0) + ); + + _pending_recv_buffs.push_with_haste(mrb); //timeout: return the managed buffer to the queue + } + return managed_recv_buffer::sptr(); + } + + size_t get_num_recv_frames(void) const {return _num_recv_frames;} + size_t get_recv_frame_size(void) const {return _recv_frame_size;} + + /******************************************************************* + * Send implementation: + * Block on the managed buffer's get call and advance the index. + ******************************************************************/ + managed_send_buffer::sptr get_send_buff(double timeout){ + if (_next_send_buff_index == _num_send_frames) _next_send_buff_index = 0; + return _msb_pool[_next_send_buff_index]->get_new(timeout, _next_send_buff_index); + } + + size_t get_num_send_frames(void) const {return _num_send_frames;} + size_t get_send_frame_size(void) const {return _send_frame_size;} + +private: + //memory management -> buffers and fifos + const size_t _recv_frame_size, _num_recv_frames; + const size_t _send_frame_size, _num_send_frames; + buffer_pool::sptr _recv_buffer_pool, _send_buffer_pool; + std::vector<boost::shared_ptr<udp_zero_copy_asio_msb> > _msb_pool; + std::vector<boost::shared_ptr<udp_zero_copy_asio_mrb> > _mrb_pool; + bounded_buffer<udp_zero_copy_asio_mrb *> _pending_recv_buffs; + size_t _next_send_buff_index; + + //socket guts + SOCKET _sock_fd; +}; + +/*********************************************************************** + * UDP zero copy make function + **********************************************************************/ +udp_zero_copy::sptr udp_zero_copy::make( + const std::string &addr, + const std::string &port, + const device_addr_t &hints +){ + return sptr(new udp_zero_copy_wsa_impl(addr, port, hints)); +} |