Hi,
I'm noticing that since the HClient/HRegionServer interface only allows
for a per-column put(), there is a lot of RPC and some lease management
overhead when writing large amounts of data. For example:
for (int i = 0; i < 10000; i++) {
Text rowKey = new Text(i+"");
long lock = client.startUpdate(rowKey);
client.put(lock, COL1, rowKey.getBytes());
client.put(lock, COL2, someValue.getBytes());
client.commit(lock);
}
This code takes my machine (using a single HMaster/HRegionServer on
local filesystem) approximately 13 seconds to execute. When i measure
the execution time within HRegionServer.put() I get total time spent in
put() < 2 seconds. So it looks like there's definately overhead in the
RPC communication and serialization/deserialization between client and
server.
To write 10000 rows, 10000 x (startUpdate=1 + #cols=2 + commit=1) =
40000 RPC operations.
What I'm thinking, and please tell me if i'm wrong or if this is already
in the works, is that if I create a row-level put() method that submits
a map of column values at once, I would reduce the 2 + (#cols) RPC
operations to one single atomic row-write RPC as well as eliminate the
small but noticeable overhead in lease creation, renewal, and cancellation.
It's not clear exactly what the performance improvement would be. The
same amount of serialization/deserilalization must occur, but YourKit
profiling tells me that the serialization overhead is negligible.
Any thoughts?
