Protection Key Faults on PA-RISC System

[Premgy Varghese]

Please refer to http://docs.huihoo.com/oracle/docs/B19306_01/server.102/b15658/appb_hpux.htm
Under the title "B.1 HP-UX Shared Memory Segments for an Oracle Instance" one can find that PA-RISC architecture has 6 unique protection keys for its shared memory segments and if the number of shared segments crosses 6, HP-UX displays protection key faults. In my system `ipcs -m` shows 20+ shared segments and my DBA is concerned over the performance degradation. Given that shmmax can be increased, can someone tell me where exactly "protection key faults" are logged? I checked in `dmesg` and syslog, but it's not there.

[A. Clay Stephenson · Accepted solution]

Bear in mind that the total number of shared identifiers which is what ipcs -m displays is not of direct concern as that is the system-wide number of shmid's. What is of real concern to you, is the number of shmid's that are attached to one process. At the process level, we can refer to them as segments. You aren't going to see protection key faults in any log just as you aren't going to see page faults in any log. You might be able to use adb to extract them from a running kernel. The two are vaguely similar in nature and result from the finite resources that are built into the CPU and ancillary devices. Based upon the Oracle data, it appears that PA-RISC architecture will allow upto 6 segments to be quickly evaluated for safety and if your SGA will not fit into 6 segments then the same checks take a bit longer because they must be brought in from slower memory.

The good news is that the fix is trivially simple; set shmmax so that your largest anticipated SGA will fit in one segment. Otherwise, when a large SGA is used (ie larger than shmmax) multiple shmgets are used to create the SGA which results in multiple segments being assigned to the SGA -- and according to your reference anything over 6 triggers the extra lookups --- so set shmmax large and then you can forget about this.

[Premgy Varghese]

That was brilliant, thank you very much.

[Premgy Varghese]

The response is quite descriptive and genuine. Thanks very much.

[rick jones]

Perhaps the prospect tool <> knows how to show pidfaults? ISTR the old PA "pi" program (ftp://ftp.cup.hp.com/dist/networking/tools/) might have known, but I don't beleive it works (or should even be run) on later PA-RISC CPUs (anything past the PA-8700 IIRC, perhaps even including it)