Fix starvation in object server with fast clients.

When an object server was handling concurrent GET or POST requests
from very fast clients, it would starve other connected clients. The
greenthreads responsible for servicing the fast clients would hog the
processor and only rarely yield to another greenthread.
The reason this happens for GET requests is found in
eventlet.greenio.GreenSocket, in the send() method. When you call
.send(data) on a GreenSocket, it immediately calls .send(data) on its
underlying real socket (socket._socketobject). If the real socket
accepts all the data, then GreenSocket.send() returns without yielding
to another greenthread. Only if the real socket failed to accept all
the data (either .send(data) < len(data) or by raising EWOULDBLOCK)
does the GreenSocket yield control.
Under most workloads, this isn't a problem. The TCP connection to
client X can only consume data so quickly, and therefore the
greenthread serving client X will frequently encounter a full socket
buffer and yield control, so no clients starve. However, when there's
a lot of contention for a single object from a large number of fast
clients (e.g. on a LAN connected w/10Gb Ethernet), then one winds up
in a situation where reading from the disk is slower than writing to
the network, and so full socket buffers become rare, and therefore so
do context switches. The end result is that many clients time out
waiting for data.
The situation for PUT requests is analogous; GreenSocket.recv() seldom
encounters EWOULDBLOCK, so greenthreads seldom yield.
This patch calls eventlet.sleep() to yield control after each chunk,
preventing any one greenthread's IO from blocking the hub for very
long.
This code has the flaw that it will greenthread-switch twice when a
send() or recv() does block, but since there isn't a way to find out
if a switch occurred or not, there's no way to avoid it. Since
greenlet switches are quite fast (faster than system calls, which the
object server does a lot of), this shouldn't have a significant
performance impact.
Change-Id: I8549adfb4a198739b80979236c27b76df607eebf
This commit is contained in:
Samuel Merritt
2012年06月01日 15:24:21 -07:00
parent 7a9c2d6ea5
commit 783f16035a

View File

@@ -90,6 +90,18 @@ class TestDiskFile(unittest.TestCase):
'o', FakeLogger()).mkstemp():
self.assert_(os.path.exists(tmpdir))
def test_iter_hook(self):
hook_call_count = [0]
def hook():
hook_call_count[0] += 1
df = self._get_data_file(fsize=65, csize=8, iter_hook=hook)
print repr(df.__dict__)
for _ in df:
pass
self.assertEquals(hook_call_count[0], 9)
def test_quarantine(self):
df = object_server.DiskFile(self.testdir, 'sda1', '0', 'a', 'c', 'o',
FakeLogger())
@@ -136,7 +148,8 @@ class TestDiskFile(unittest.TestCase):
self.assert_('-' in os.path.basename(double_uuid_path))
def _get_data_file(self, invalid_type=None, obj_name='o',
fsize=1024, csize=8, extension='.data', ts=None):
fsize=1024, csize=8, extension='.data', ts=None,
iter_hook=None):
'''returns a DiskFile'''
df = object_server.DiskFile(self.testdir, 'sda1', '0', 'a', 'c',
obj_name, FakeLogger())
@@ -168,7 +181,8 @@ class TestDiskFile(unittest.TestCase):
df = object_server.DiskFile(self.testdir, 'sda1', '0', 'a', 'c',
obj_name, FakeLogger(),
keep_data_fp=True, disk_chunk_size=csize)
keep_data_fp=True, disk_chunk_size=csize,
iter_hook=iter_hook)
if invalid_type == 'Zero-Byte':
os.remove(df.data_file)
fp = open(df.data_file, 'w')
Reference in New Issue
openstack/swift
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.

The note is not visible to the blocked user.