Resolving ORA-27090: Unable to reserve kernel resources for asynchronous disk I/O in Oracle 12c

I have been running into some problems recently with 12cR2 databases and Kevin Clossons’ SLOB tool.

The SLOB script allows for the concurrent loading of multiple schemas, and if you are loading a large amount of data, being able to load concurrently is a significant time saver.

With LOAD_PARALLEL_DEGREE set to 8, I got the following error:

ORA-27090: Unable to reserve kernel resources for asynchronous disk I/O
Linux-x86_64 Error: 11: Resource temporarily unavailable
Additional information: 3
Additional information: 128
Additional information: 140728056780720

These servers were new Dell R630s with plenty of horsepower, so the idea that just 8 parallel threads would cause this type of a failure was puzzling.

Further investigation of the trace file showed that the problem occured on the index shrink command:

ERROR at line 1:
ORA-12801: error signaled in parallel query server P13L, instance
sio04-mgmt.asp.lab.mcl:slob4 (4)

After some time investigating, it seems Oracle 12c has a much higher target for PARALLEL_MAX_SERVERS and PARALLEL_SERVERS_TARGET.  In my case, PARALLEL_MAX_SERVERS had defaulted to 2240.

Since the SLOB data load uses the parallel query option, Oracle was spawning thousands of slave processes all trying to issue ASYNC IO.

So I set the numbers to what I considered more reasonable:

SQL> alter system set parallel_max_servers=400 sid='*';

System altered.

SQL> alter system set parallel_min_servers=40 sid='*';

System altered.

SQL> alter system set parallel_servers_target=400 sid='*';

System altered.

Now SLOB was able to load data with eight concurrent processes.

EMC Unity Storage Performance testing with Oracle ASM and SLOB

I’ve been testing the new EMC Unity 600F all-flash storage array with an Oracle database to determine the impact of storage compression on Oracle I/O performance.  To do this I have been using Kevin Closson’s SLOB tool.

SLOB is an excellent tool for testing I/O performance as seen by Oracle.  Unlike Swingbench which mimics a real application and therefore spends much of its execution cycle on the server CPU, SLOB concentrates exclusively on generating I/O from the Oracle stack.

Conversely, don’t expect SLOB to generate meaningful data to test ranking or sorting operations inside of Oracle.  SLOB generates entirely synthetic data that is meaningless from an application standpoint.

The following posts covers using SLOB to test I/O performance, and what was learned from the testing against the Unity 600F all-flash array.

Continue reading