OFI libfabric原理及应用解析

Agenda 目录/议题

编译通信软件
硬件和软件带来的挑战
为什么需要libfabric
libfabric架构
API分组
socket应用 VS libfabric应用区别
GPU数据传输示例

编译通信软件

可靠面向连接的TCP和无连接的数据报UDP协议
高性能计算HPC或人工智能AI

软硬件复杂性带来的挑战

上千个节点的集群, 不同的网络类型(以太网, IB, 光纤等), 熊猫博士提到的CPU/GPU/XPU/IPU等等
xelink, nvlink, gpu
软件库, 如nccl,intel mpi等
所以咱们需要一个通用的通信库来桥接这些复杂的软硬件资源

libfabric来解决上面的问题

统一API, 让程序员更轻松
高性能和高可扩展性
核心组件: 众多网卡提供者的库, 核心服务, 测试程序等

为什么需要libfabric

承上启下, 桥接底层复杂多样的网络(socket, rdma, gpu, 共享内存等)和上层MPI, CCL, 共享内存等应用

socket编程与libfabric编程对比

语义类似, 如获取信息, bind, connect等, 但是libfabric底层支持多种网络类型

支持GPU通信

gpu通信示例, 支持intel gpu, dpu, 或者其他供应商

架构与四种服务

控制类: 发现底层设备, 属性, 能力等
通信接口: 建立连接, 初始资源等
数据传输: 发送和接收数据
完成服务: 报告发送或接收状态

libfabric API分组

整合底层提供者和上层开发者使用统一的API编程, 就像熊猫博士说的那样, 方便了提供者更方便的提供插件, 也方便了上层应用开发者

tcp socket 代码截图

socket收发数据示例

server端启动:./example_socket客户端连接和发送:./example_socket 192.168.5.6Hello server this is client.

example_socket.c 源码

#include <stdio.h>
#include <string.h>
#include <sys/socket.h>
#include <arpa/inet.h>char *dst_addr = NULL;int main(int argc, char *argv[])
{int socket_desc, client_sock, client_size;struct sockaddr_in server_addr, client_addr;char server_message[2000], client_message[2000];dst_addr = inet_addr(argv[1]);// Clean buffers:memset(server_message, '\0', sizeof(server_message));memset(client_message, '\0', sizeof(client_message));// Create socket:socket_desc = socket(AF_INET, SOCK_STREAM, 0);if(socket_desc < 0){printf("Error while creating socket\n");return -1;}printf("Socket created successfully\n");// Set port and IP:server_addr.sin_family = AF_INET;server_addr.sin_port = "43192";server_addr.sin_addr.s_addr = inet_addr("127.0.0.1");if (!dst_addr) {// Bind to the set port and IP:if(bind(socket_desc, (struct sockaddr*)&server_addr, sizeof(server_addr))<0){printf("Couldn't bind to the port\n");return -1;}printf("Binding complete\n");// Listen for clients:if(listen(socket_desc, 1) < 0){printf("Error while listening\n");return -1;}printf("Listening for incoming connections...\n");// Accept an incoming connection:client_size = sizeof(client_addr);client_sock = accept(socket_desc, (struct sockaddr*)&client_addr, &client_size);if (client_sock < 0){printf("Can't accept\n");return -1;}printf("Client connected at IP: %s and port: %i\n", inet_ntoa(client_addr.sin_addr), ntohs(client_addr.sin_port));// Receive client's message:if (recv(client_sock, client_message, sizeof(client_message), 0) < 0){printf("Couldn't receive\n");return -1;}printf("Msg from client: %s\n", client_message);// Respond to client:strcpy(server_message, "This is the server's message.");if (send(client_sock, server_message, strlen(server_message), 0) < 0){printf("Can't send\n");return -1;}}if (dst_addr) {// Send connection request to server:if(connect(socket_desc, (struct sockaddr*)&server_addr, sizeof(server_addr)) < 0){printf("Unable to connect\n");return -1;}printf("Connected with server successfully\n");// Get input from the user:printf("Enter message: ");gets(client_message);// Send the message to server:if(send(socket_desc, client_message, strlen(client_message), 0) < 0){printf("Unable to send message\n");return -1;}// Receive the server's response:if(recv(socket_desc, server_message, sizeof(server_message), 0) < 0){printf("Error while receiving server's msg\n");return -1;}printf("Server's response: %s\n",server_message);}// Close the client socket:close(client_sock);// Closing the server socket:close(socket_desc);return 0;
}

libfabric收发数据示例截图

服务端启动: ./example_msg客户端连接发送数据: ./example_msg 192.168.5.6
客户端连接发送数据: ./example_msg 192.168.5.6

参考代码

https://github.com/ssbandjl/libfabric/blob/main/fabtests/functional/example_msg.c
/*** This software is available to you under the BSD license* below:**     Redistribution and use in source and binary forms, with or*     without modification, are permitted provided that the following*     conditions are met:**      - Redistributions of source code must retain the above*        copyright notice, this list of conditions and the following*        disclaimer.**      - Redistributions in binary form must reproduce the above*        copyright notice, this list of conditions and the following*        disclaimer in the documentation and/or other materials*        provided with the distribution.** THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND* NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS* BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN* ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN* CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE* SOFTWARE.*/#include <stdio.h>
#include <stdlib.h>
#include <getopt.h>
#include <netinet/in.h>
#include <arpa/inet.h>
#include <netdb.h>
#include <rdma/fabric.h>
#include <rdma/fi_domain.h>
#include <rdma/fi_endpoint.h>
#include <rdma/fi_cm.h>
#include <shared.h>//Build with
//gcc -o example_msg example_msg.c -L<path to libfabric lib> -I<path to libfabric include> -lfabric
//gcc -o example_msg example_msg.c -L/home/xb/project/libfabric/libfabric/build/lib -I/home/xb/project/libfabric/libfabric/build/include -I/home/xb/project/libfabric/libfabric/build/include -lfabric#define BUF_SIZE 64char *dst_addr = NULL;
char *port = "9228";
struct fi_info *hints, *info, *fi_pep;
struct fid_fabric *fabric = NULL;
struct fid_domain *domain = NULL;
struct fid_ep *ep = NULL;
struct fid_pep *pep = NULL;
struct fid_cq *cq = NULL;
struct fid_eq *eq = NULL;
struct fi_cq_attr cq_attr = {0};
struct fi_eq_attr eq_attr = {.wait_obj = FI_WAIT_UNSPEC
};
//const struct sockaddr_in *sin;
char str_addr[INET_ADDRSTRLEN];
int ret;
char buffer[BUF_SIZE];
fi_addr_t fi_addr = FI_ADDR_UNSPEC;
struct fi_eq_cm_entry entry;
uint32_t event;
ssize_t rd;/* Initializes all basic OFI resources to allow for a server/client to exchange a message */
static int start_client(void)
{ret = fi_getinfo(FI_VERSION(1,9), dst_addr, port, dst_addr ? 0 : FI_SOURCE,hints, &info);if (ret) {printf("fi_getinfo: %d\n", ret);return ret;}ret = fi_fabric(info->fabric_attr, &fabric, NULL);if (ret) {printf("fi_fabric: %d\n", ret);return ret;}ret = fi_eq_open(fabric, &eq_attr, &eq, NULL);if (ret) {printf("fi_eq_open: %d\n", ret);return ret;}ret = fi_domain(fabric, info, &domain, NULL);if (ret) {printf("fi_domain: %d\n", ret);return ret;}/* Initialize our completion queue. Completion queues are used to report events associated* with data transfers. In this example, we use one CQ that tracks sends and receives, but* often times there will be separate CQs for sends and receives.*/cq_attr.size = 128;cq_attr.format = FI_CQ_FORMAT_MSG;ret = fi_cq_open(domain, &cq_attr, &cq, NULL);if (ret) {printf("fi_cq_open error (%d)\n", ret);return ret;}/* Bind our CQ to our endpoint to track any sends and receives that come in or out on that endpoint.* A CQ can be bound to multiple endpoints but one EP can only have one send CQ and one receive CQ* (which can be the same CQ).*/ret = fi_endpoint(domain, info, &ep, NULL);if (ret) {printf("fi_endpoint: %d\n", ret);return ret;}ret = fi_ep_bind(ep, &cq->fid, FI_SEND | FI_RECV);if (ret) {printf("fi_ep_bind cq error (%d)\n", ret);return ret;}ret = fi_ep_bind((ep), &(eq)->fid, 0);if (ret) {printf("fi_ep_bind: %d\n", ret);return ret;}ret = fi_enable(ep);if (ret) {printf("fi_enable: %d\n", ret);return ret;}ret = fi_connect(ep, info->dest_addr, NULL, 0);if (ret) {printf("fi_connect: %d\n", ret);return ret;}rd = fi_eq_sread(eq, &event, &entry, sizeof(entry), -1, 0);if (rd != sizeof(entry)) {ret = (int) rd;printf("fi_eq_sread: %d\n", ret);return ret;}return 0;
}static int start_server(void)
{const struct sockaddr_in *sin;/* The first OFI call to happen for initialization is fi_getinfo which queries libfabric* and returns any appropriate providers that fulfill the hints requirements. Any applicable* providers will be returned as a list of fi_info structs (&info). Any info can be selected.* In this test we select the first fi_info struct. Assuming all hints were set appropriately,* the first fi_info should be most appropriate.* The flag FI_SOURCE is set for the server to indicate that the address/port refer to source* information. This is not set for the client because the fields refer to the server, not* the caller (client). *//* 初始化时发生的第一个 OFI 调用是 fi_getinfo，它查询 libfabric 并返回满足提示要求的任何适当的提供程序。 任何适用的提供程序都将作为 fi_info 结构 (&info) 列表返回。 可以选择任何信息。 在此测试中，我们选择第一个 fi_info 结构。 假设所有提示均已正确设置，第一个 fi_info 应该是最合适的。 为服务器设置标志FI_SOURCE以指示地址/端口引用源信息。 这不是为客户端设置的，因为这些字段引用服务器，而不是调用者（客户端） */ret = fi_getinfo(FI_VERSION(1,9), dst_addr, port, dst_addr ? 0 : FI_SOURCE,hints, &fi_pep);if (ret) {printf("fi_getinfo error (%d)\n", ret);return ret;}/* Initialize our fabric. The fabric network represents a collection of hardware and software* resources that access a single physical or virtual network. All network ports on a system* that can communicate with each other through their attached networks belong to the same fabric.*/ret = fi_fabric(fi_pep->fabric_attr, &fabric, NULL); // 打开fabric, 初始化任何资源前需要打开fabricif (ret) {printf("fi_fabric error (%d)\n", ret);return ret;}/* Initialize our endpoint. Endpoints are transport level communication portals which are used to* initiate and drive communication. There are three main types of endpoints:* FI_EP_MSG - connected, reliable* FI_EP_RDM - unconnected, reliable* FI_EP_DGRAM - unconnected, unreliable* The type of endpoint will be requested in hints/fi_getinfo. Different providers support different* types of endpoints. TCP supports only FI_EP_MSG but when used with RxM, can support FI_EP_RDM.* In this application, we requested TCP and FI_EP_MSG.*/ret = fi_eq_open(fabric, &eq_attr, &eq, NULL); // 打开事件队列EQ, 一般用于建连, 收发数据产生的事件if (ret) {printf("fi_eq_open: %d\n", ret);return ret;}ret = fi_passive_ep(fabric, fi_pep, &pep, NULL); // 打开被动端点, 常用与服务端监听端口, 支持多个客户端domain连接进来if (ret) {printf("fi_passive_ep: %d\n", ret);return ret;}ret = fi_pep_bind(pep, &eq->fid, 0); // 为端点绑定事件队列if (ret) {printf("fi_pep_bind %d", ret);return ret;}ret = fi_listen(pep); // 监听端点, 等待客户端连接请求if (ret) {printf("fi_listen %d", ret);return ret;}return 0;
}static int complete_connection(void)
{rd = fi_eq_sread(eq, &event, &entry, sizeof(entry), -1, 0); // 等待读取客户端触发的服务端事件, 读取事件, 推动进展(驱动程序运转)if (rd != sizeof entry) {ret = (int) rd;printf("fi_eq_sread: %d", ret);if (ret)goto err;}ret = fi_domain(fabric, info, &domain, NULL); // domain域用于将资源分组, 可基于域来做管理if (ret) {printf("fi_domain: %d\n", ret);return ret;}ret = fi_domain_bind(domain, &eq->fid, 0);if (ret) {printf("fi_domain_bind: %d\n", ret);return ret;}/* * Initialize our completion queue. Completion queues are used to report events associated* with data transfers. In this example, we use one CQ that tracks sends and receives, but* often times there will be separate CQs for sends and receives.*/cq_attr.size = 128;cq_attr.format = FI_CQ_FORMAT_MSG;ret = fi_cq_open(domain, &cq_attr, &cq, NULL);if (ret) {printf("fi_cq_open error (%d)\n", ret);return ret;}/* Bind our CQ to our endpoint to track any sends and receives that * come in or out on that endpoint. A CQ can be bound to multiple* endpoints but one EP can only have one send CQ and one receive CQ* (which can be the same CQ).*/ret = fi_endpoint(domain, info, &ep, NULL); // 用于客户端, 主动端点, 发起建连if (ret) {printf("fi_endpoint: %d\n", ret);return ret;}ret = fi_ep_bind(ep, &cq->fid, FI_SEND | FI_RECV);if (ret) {printf("fi_ep_bind cq error (%d)\n", ret);return ret;}ret = fi_ep_bind((ep), &(eq)->fid, 0);if (ret) {printf("fi_ep_bind: %d\n", ret);return ret;}ret = fi_enable(ep);if (ret) {printf("fi_enable: %d", ret);return ret;}ret = fi_accept(ep, NULL, 0);if (ret) {printf("fi_accept: %d\n", ret);return ret;}rd = fi_eq_sread(eq, &event, &entry, sizeof(entry), -1, 0);if (rd != sizeof(entry)) {ret = (int) rd;printf("fi_eq_read: %d\n", ret);return ret;}return 0;err:if (info)fi_reject(pep, info->handle, NULL, 0);return ret;}static void cleanup(void)
{int ret;/* All OFI resources are cleaned up using the same fi_close(fid) call. */if (ep) {ret = fi_close(&ep->fid);if (ret)printf("warning: error closing EP (%d)\n", ret);}if (pep) {ret = fi_close(&pep->fid);if (ret)printf("warning: error closing PEP (%d)\n", ret);}ret = fi_close(&cq->fid);if (ret)printf("warning: error closing CQ (%d)\n", ret);ret = fi_close(&domain->fid);if (ret)printf("warning: error closing domain (%d)\n", ret);ret = fi_close(&eq->fid);if (ret)printf("warning: error closing EQ (%d)\n", ret);ret = fi_close(&fabric->fid);if (ret)printf("warning: error closing fabric (%d)\n", ret);if (info)fi_freeinfo(info);if (fi_pep)fi_freeinfo(fi_pep);
}/* Post a receive buffer. This call does not ensure a message has been received, just* that a buffer has been passed to OFI for the next message the provider receives.* Receives may be directed or undirected using the address parameter. Here, we* pass in the fi_addr but note that the server has not inserted the client's* address into its AV, so the address is still FI_ADDR_UNSPEC, indicating that* this buffer may receive incoming data from any address. An application may* set this to a real fi_addr if the buffer should only receive data from a certain* peer.* When posting a buffer, if the provider is not ready to process messages (because* of connection initialization for example), it may return -FI_EAGAIN. This does* not indicate an error, but rather that the application should try again later.* This is why we almost always wrap sends and receives in a do/while. Some providers* may need the application to drive progress in order to get out of the -FI_EAGAIN* loop. To drive progress, the application needs to call fi_cq_read (not necessarily* reading any completion entries).*/
static int post_recv(void)
{int ret;do {ret = fi_recv(ep, buffer, BUF_SIZE, NULL, fi_addr, NULL);if (ret && ret != -FI_EAGAIN) {printf("error posting recv buffer (%d\n", ret);return ret;}if (ret == -FI_EAGAIN)(void) fi_cq_read(cq, NULL, 0);} while (ret);return 0;
}/* Post a send buffer. This call does not ensure a message has been sent, just that* a buffer has been submitted to OFI to be sent. Unlike a receive buffer, a send* needs a valid fi_addr as input to tell the provider where to send the message.* Similar to the receive buffer posting porcess, when posting a send buffer, if the* provider is not ready to process messages, it may return -FI_EAGAIN. This does not* indicate an error, but rather that the application should try again later. Just like* the receive, we drive progress with fi_cq_read if this is the case.*/
static int post_send(void)
{char *msg = "Hello, server! I am the client you've been waiting for!\0";int ret;(void) snprintf(buffer, BUF_SIZE, "%s", msg);do {ret = fi_send(ep, buffer, BUF_SIZE, NULL, fi_addr, NULL);if (ret && ret != -FI_EAGAIN) {printf("error posting send buffer (%d)\n", ret);return ret;}if (ret == -FI_EAGAIN)(void) fi_cq_read(cq, NULL, 0);} while (ret);return 0;
}/* Wait for the message to be sent/received using the CQ. fi_cq_read not only drives progress* but also returns any completed events to notify the application that it can reuse* the send/recv buffer. The returned completion entry will have fields set to let the application* know what operation completed. Not all fields will be valid. The fields set will be indicated* by the cq format (when creating the CQ). In this example, we use FI_CQ_FORMAT_MSG in order to* use the flags field.*/
static int wait_cq(void)
{struct fi_cq_err_entry comp;int ret;do {ret = fi_cq_read(cq, &comp, 1);if (ret < 0 && ret != -FI_EAGAIN) {printf("error reading cq (%d)\n", ret);return ret;}} while (ret != 1);if (comp.flags & FI_RECV)printf("I received a message!\n");else if (comp.flags & FI_SEND)printf("My message got sent!\n");return 0;
}static int run(void)
{int ret;if (dst_addr) {printf("Client: send to server %s\n", dst_addr);ret = post_send();if (ret)return ret;ret = wait_cq();if (ret)return ret;} else {printf("Server: post buffer and wait for message from client\n");ret = post_recv();if (ret)return ret;ret = wait_cq();if (ret)return ret;printf("This is the message I received: %s\n", buffer);}return 1;
}int main(int argc, char **argv)
{int ret;/* Hints are used to request support for specific features from a provider */hints = fi_allocinfo(); // if (!hints)return EXIT_FAILURE;/* Server run with no args, client has server's address as an argument */dst_addr = argv[1];//Set anything in hints that the application needs/* Request FI_EP_MSG (reliable datagram) endpoint which will allow us* to reliably send messages to peers without having to listen/connect/accept. */hints->ep_attr->type = FI_EP_MSG; // 可靠数据报端点, 类似socket, 但无须执行listen/connect/accept/* Request basic messaging capabilities from the provider (no tag matching,* no RMA, no atomic operations) */hints->caps = FI_MSG;/* Specifically request the tcp provider for the simple test */// hints->fabric_attr->prov_name = "tcp"; // 类似socket的, 面向连接的消息类型端点hints->fabric_attr->prov_name = "ofi_rxm;verbs";/* Specifically request SOCKADDR_IN address format to simplify addressing for test */hints->addr_format = FI_SOCKADDR_IN;/* Default to FI_DELIVERY_COMPLETE which will make sure completions do not get generated* until our message arrives at the destination. Otherwise, the client might get a completion* and exit before the server receives the message. This is to make the test simpler *//* 默认为 FI_DELIVERY_COMPLETE，这将确保在我们的消息到达目的地之前不会生成完成(等待)。 否则，客户端可能会在服务器收到消息之前完成并退出。 这是为了让测试更简单 */hints->tx_attr->op_flags = FI_DELIVERY_COMPLETE;//Done setting hintsif (!dst_addr) {ret = start_server();if (ret) {goto out;return ret;}}ret = dst_addr ? start_client() : complete_connection();if (ret) {goto out;return ret;}ret = run();
out:cleanup();return ret;
}

socket vs libfabric消息类型示意图(两者都可完成建连和消息收发)

GPU数据传输示例

左边是内存直接访问DMA ibv verbs示例, 右边是DMA libfabric统一API的语义的示例

verbs
服务端: ./rdmabw-xe -m host
客户端: ./rdmabw-xe -m host -S 1 -t write 192.168.5.6  #都是用主机内存, 完成1字节的远程内存写操作

主机对主机GPU libfabric 内存直接访问DMA示例

libfabric, host -> host DMA
服务端: ./fi-rdmabw-xe -m host
客户端: ./fi-rdmabw-xe -m host -S 1 -t write 192.168.5.6  #都是用主机内存, 完成1字节的远程内存写操作
verbs 代码相对位置: fabtests/component/dmabuf-rdma/rdmabw-xe.c

主机发给GPU设备内存直接访问DMA libfabric示例

libfabric GPU设备 -> host
服务端: ./fi-rdmabw-xe -m device #使用GPU设备的内存
客户端: ./fi-rdmabw-xe -m host -S 1 -t write 192.168.5.6  #用主机内存, 完成1字节的远程内存写操作
libfabric 代码相对位置: fabtests/component/dmabuf-rdma/fi-rdmabw-xe.c