GBL_LOST_MI_TRACE_BUFFERS warnings : How to fix ??

Error :  Measurement Buffers Lost see metric GBL_LOST_MI_TRACE_BUFFERS warnings HP UNIX.

Reason :  This warnings are frequently expect on high spec busy server. The midaemon process gets its raw information from “event traces” that are sent from the kernel instrumentation (KI) that is an integral part of HPUX. This instrumentation is written into virtually every system call in the kernel. Whenever a process enters or leaves a system call, one or more event traces is sent “over the fence” to the midaemon process. Midaemon has two threads…a reader thread and a writer thread. The reader thread accepts these event traces, while the writer thread “massages” them somewhat, and then writes the pertinent information out to the midaemon "shared memory database" or SMD, a table that is allocated in the shared memory space. Now, think of a large funnel…at the top or large end of the funnel there are multiple processors pouring in these event traces, while at the bottom or small end of the funnel there is one processor running the midaemon reader thread pulling the event traces out. On a busy, good-sized system, say 16 processors or more, it is possible that this logical “funnel” can overflow, resulting in the GBL_LOST_MI_TRACE_BUFFER messages.

How to fix ??

What can be done to reduce the or eliminate:-

For Glance version :- C.04.XX.XXX

1. Contact support and request the C.04.73.115 hotfix for OVPA.

2. Set the "-bufsets" and "-skipbuf" parameters in the midaemon startup command appropriately. System adminstators could consider these values as appropriate settings:
No. Of CPUs midaemon parameters
8 -bufsets 16 -skipbuf 8 -smdvss 256M
16 -bufsets 24 -skipbuf 12 -smdvss 256M
32 -bufsets 32 -skipbuf 16 -smdvss 512M
64 -bufsets 32 -skipbuf 16 -smdvss 512M
Appropriate values for larger numbers of CPUs, e.g. 128 and 256 are still being investigated. Setting the above parameters requires that the "maxdsize" kernel parameter be set to at least 2 GB.

3. Utilize the undocumented "-no_fileio_traces" parameter in the midaemon startup command string.

This is best done by editing the /etc/rc.config.d/ovpa file. In order to set the appropriate specifications for a 64 CPU system with "-no_fileio_traces", for example, would entail adding
The following two lines to that file:

MIPARMS = "-p -no_fileio_traces -bufsets 32 -skipbuf 16 -smdvss 512M"
export MIPARMS

For Glance version :- C.05.XX.XXX

To activate logical IO metric collection and display, use the following procedures:
If the Performance Agent (PA) and Glanceplus are both installed, follow these seven steps:

1. Change directory to /etc/rc.config.d
2. Edit the "ovpa" file using vi or the editor of choice
3. Find the line immediately preceding MWA_START = 1
4. Add the following two statements and save the change:
MIPARMS="-fileio_traces -p”
export MIPARMS
5. Run the commands: “ovpa stop” and “midaemon -T”. Check the midaemon status with “perfstat -p” to make sure it has terminated and is ready to use the new configuration.
6. Execute the command: “ovpa start”
7. Start Glance, go to the "IO by Disk" screen using the “d” key, and check to ensure that Glance is now populating the logl rds and logl wrts (logical IO) metrics.

If only Glanceplus is installed on the system, i.e. the Performance Agent is not installed, then replace the above steps with the following:

1. Stop all executions of glance and midaemon. Use "ps -ef | grep midaemon" to ensure that there is no active midaemon process.
2. Start a new midaemon process using the command "midaemon -p -fileio_traces".
3. Execute step #7 above.

Post a Comment