I am seeing some significant performance differences between OracleBulkCopy (ODP.NET) and SQL*Loader when the Oracle server is on another machine.
I have a very basic table in Oracle with three columns (one BINARY_FLOAT, two NUMBER(18,0)). There are no primary keys, indexes, triggers, etc. It is used as a staging table to get bulk data into the DB.
SQL*Loader takes about 27 seconds to load 4.5 million rows into the table.
OracleBulkCopy takes about 10 minutes to load just 1 million rows.
OracleBulkCopy, according to the documentation, *”…uses a direct path load approach, which is similar to, but not the same as Oracle SQL*Loader.”* It might not be up there with SQL*Loader in terms of performance, but this difference is ridiculous.
After some basic analysis of the network traffic, I found the OracleBulkCopy was sending and receiving a huge number of small packets. I used Wireshark to compare the packets for each and found some interesting results.
SQL*Loader – after the initial connection handshaking – sends a series of 8 kilobyte packets (TNS protocol) and receives 60 byte ACKs in response.
OracleBulkCopy sends a series of 102 byte packets (TNS protocol) and receives a 133 byte packet (TNS protocol) in response. What the…!? It is like it is sending one row at a time!
With the OracleBulkCopy class, I am using a batch size of 100,000 and am using a custom IDataReader to read from a data file.
So, my questions are:
-
Has anyone ever seen this behaviour?
-
Does OracleBulkCopy actually perform as a bulk loading tool?
-
Is there something I need to configure to get it to work properly? (client/server settings, etc)
Any help is much appreciated.
I received a response from Alex Keh, Oracle Product Manager, that Oracle “has noticed this issue as well. We’re evaluating how to fix this bug.”
So in answer to my own questions, I guess OracleBulkCopy does not perform well as a bulk loading tool. At least not at the moment.
I will be wrapping SQL*Loader as an alternate solution as there is no ETA on the bug fix.