tornado并发介绍
tornado处理请求是单线程的,如果所有操作都在这个线程做,qps就受限于你的业务复杂性。业务比较复杂时可以起协程去做一些耗时的操作,主线程处理一些简单操作。需要注意的是起的协程不宜过多,协程过多,上下文切换的开销就过大,得不偿失;如果真有这么多,那就应该起多进程去处理这些请求。
协程加多线程实现异步
#耗时的事务函数
@tornado.gen.coroutine
def embedding(sentence):
do something
raise Return(your result)
#线程池
thread_pool = ThreadPoolExecutor(4)
#handler的请求处理函数
@tornado.gen.coroutine
def get(self):
query = self.get_argument('sentence', default='', strip=True)
result = thread_pool.submit(embedding, ([query]))
s = yield result
self.write("%s" % (s))
@tornado.gen.coroutine表示可以协程去运行这个函数,调用协程的函数必须是可以协程调用的,thread_pool.submit表示提交一个任务到线程池,会立即返回一个future对象,yield result会等待这个future对象完成,但是由于是协程的方式执行,这个地方并不会阻塞,主线程会保存上下文,处理其他请求,等result这个future对象执行完毕再切换回来执行后面的代码,感兴趣可以去看future实现。
多进程误区
- centOS上http_server应该是bind,不要listen,参照官方文档https://www.tornadoweb.org/en/stable/httpserver.html
- autoreload=False debug=False,多线程的必须这样设置,debug=True会让autoreload=True
- 由于多进程是fork的,变量的初始化最好在fork之后,例如显卡资源先申请了,fork之后并不能多个进程共用。
完整代码如下:
#encoding=utf-8
import time
import tornado.web
import tornado.ioloop
import sys
import random
import tensorflow as tf
from bert_serving.client import ConcurrentBertClient
from tornado.concurrent import run_on_executor
from concurrent.futures import ThreadPoolExecutor
from tornado.gen import Return
bc = None
sess = None
output = None
thread_pool = None
input_op = None
#load_bert_embedding()
def init():
global sess, output, input_op, thread_pool, bc
saver = tf.train.import_meta_graph('cnn.final.meta')
sess = tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True)))
saver.restore(sess, tf.train.latest_checkpoint("test"))
graph = tf.get_default_graph()
input_op = graph.get_operation_by_name('Reshape_1').outputs[0]
output = graph.get_operation_by_name('BiasAdd_1').outputs[0]
bc = ConcurrentBertClient(port=5555, port_out=5556)
thread_pool = ThreadPoolExecutor(4)
@tornado.gen.coroutine
def embedding(sentence):
bert_embedding = bc.encode(sentence)
feed_dict = {input_op.name: bert_embedding}
emb = sess.run(output, feed_dict=feed_dict)
raise Return(" ".join(map(str, list(emb[0]))))
class embeddingHandler(tornado.web.RequestHandler):
@tornado.gen.coroutine
def get(self):
query = self.get_argument('sentence', default='', strip=True)
if thread_pool is None:
init()
result = thread_pool.submit(embedding, ([query]))
s = yield result
s = yield s
self.write("%s" % (s))
self.finish()
if __name__ == "__main__":
app = tornado.web.Application([(r'/get_embedding', embeddingHandler)], autoreload=False, debug=False)
sockets = tornado.netutil.bind_sockets(8888)
tornado.process.fork_processes(3)
http_server = tornado.httpserver.HTTPServer(app)
http_server.add_sockets(sockets)
tornado.ioloop.IOLoop.current().start()