The HBase client uses 1 socket per regionserver for communication. This is good for socket control but potentially bad for latency. How bad? I did a simple YCSB test that had this config:
I ran this with 1 and 10 threads. The summary is as so:
[SCAN] Operations 1000
[SCAN] AverageLatency(ms) 35.871
[SCAN] Operations 1000
[SCAN] AverageLatency(ms) 228.576
We are taking a 6.5x latency hit in our client. But why?
First step was to move the deserialization out of the Connection thread, this seemed like it could have a big win, an analog change on the server side got a 20% performance improvement (already commited as
HBASE-2941). I did this and got about a 20% improvement again, with that 228ms number going to about 190 ms.
So I then wrote a high performance nanosecond resolution tracing utility. Clients can flag an API call, and we get tracing and numbers through the client pipeline. What I found is that a lot of time is being spent in receiving the response from the network. The code block is like so:
LOG.debug(getName() + " got value #" + id);
Call call = calls.get(id);
size -= 4; // 4 byte off for id because we already read it.
ByteBuffer buf = ByteBuffer.allocate(size);
IOUtils.readFully(in, buf.array(), buf.arrayOffset(), size);
NanoProfiler.split(id, "setResponse", "Data size: " + size);
I came up with some numbers:
11726 (receiveResponse) split: 64991689 overall: 133562895 Data size: 4288937
12163 (receiveResponse) split: 32743954 overall: 103787420 Data size: 1606273
12561 (receiveResponse) split: 3517940 overall: 83346740 Data size: 4
12136 (receiveResponse) split: 64448701 overall: 203872573 Data size: 3570569
The first number is the internal counter for keeping requests unique from HTable on down. The numbers are in ns, the data size is in bytes.
Doing some simple calculations, we see for the first line we were reading at about 31 MB/sec. The second one is even worse. Other calls are like:
26 (receiveResponse) split: 7985400 overall: 21546226 Data size: 850429
which is 107 MB/sec which is pretty close to the maximum of gige. In my set up, the ycsb client ran on the master node and HAD to use network to talk to regionservers.
Even at full line rate, we could still see unacceptable hold ups of unrelated calls that just happen to need to talk to the same regionserver.
This issue is about these findings, what to do, how to improve.