During a PUT of an object, the proxy instanciates one Putter per
object-server that will store data (either the full object or a
fragment, depending on the storage policy). Each Putter is owning a
Queue that will be used to bufferize data chunks before they are
written to the socket connected to the object-server. The chunks are
moved from the queue to the socket by a greenthread. There is one
greenthread per Putter. If the client is uploading faster than the
object-servers can manage, the Queue could grow and consume a lot of
memory. To avoid that, the queue is bounded (default: 10). Having a
bounded queue also allows to ensure that all object-servers will get
the data at the same rate because if one queue is full, the
greenthread reading from the client socket will block when trying to
write to the queue. So the global rate is the one of the slowest
object-server.
The thing is, every operating system manages socket buffers for incoming
and outgoing data. Concerning the send buffer, the behavior is such that
if the buffer is full, a call to write() will block, otherwise the call
will return immediately. It behaves a lot like the Putter's Queue,
except that the size of the buffer is dynamic so it adapts itself to the
speed of the receiver.
Thus, managing a queue in addition to the socket send buffer is a
duplicate queueing/buffering that provides no interest but is, as shown
by profiling and benchmarks, very CPU costly.
This patch removes the queuing mecanism. Instead, the greenthread
reading data from the client will directly write to the socket. If an
object-server is getting slow, the buffer will fulfill, blocking the
reader greenthread. Benchmark shows a CPU consumption reduction of more
than 30% will the observed rate for an upload is increasing by about
45%.
Change-Id: Icf8f800cb25096f93d3faa1e6ec091eb29500758
2f1111a436