Operating System - OpenVMS
1748272 Members
4330 Online
108760 Solutions
New Discussion юеВ

Re: Speed of writing dump file and DUMPSTYLE settings

 
SOLVED
Go to solution
Cass Witkowski
Trusted Contributor

Speed of writing dump file and DUMPSTYLE settings

I would like to get your input for DUMPSTYLE settings.

Now that disks are larger what is the consensus of writing a selective dump versus a full dump? I think selective dumps may not write the contents of reserved memory. Is this important?

What is the typicial speed MB/S for writing to a dump file. Let assume a MSA1000 with 15K rpm disk.

I was looking at the sizing algorithm for sizing the SYSDUMP.DMP file and it looks like the size recommended for a compresses dump is 2/3 the full dump size. If I have a DS10, AS4100, or GS140 do I get a slower dump trying to compress the data to dump versus writing more data but requiring less CPU power?

Thanks

Cass
5 REPLIES 5
John Travell
Valued Contributor

Re: Speed of writing dump file and DUMPSTYLE settings

True disks are larger, but not that much faster. My understanding is that any Alpha that is EV5 or later can compress 3 pages to 2 faster than any I/O channel can write 2 pages.
Presuming that this is still true, even with 4Gb FC, you have to judge whether your priority is minimal outage time in the event of a crash, or completeness of the dump file.
You CAN have a compressed selective dump MUCH smaller than 2/3 the size of memory. Remember what goes into a selective dump, and the order. System page table, Global page table, the important parts of System space, then global pages, finally complete processes.
Things that do NOT get saved include XQP cache lines and some other objects that are rarely needed for analysis.
As each logical memory block is written, the bugcheck code compares the size of the next logical memory block with the remaining space in the file, and saves it if there is enough room. Each process is a complete logical memory block. If there is not enough space remaining that process does not get saved, until there are no processes left that are small enough to fit into the last free chunk of the dumpfile.
If you have a small file, this could be a lot of processes not saved, a much shorter dumpfile write time, and an earlier start to the reboot. On the other hand, an uncompressed 32Gb dumpfile could take a long time to write...
You choose what is best for your systems.

I do not have any figures on dumpfile write speed. Is this hypothetical dump being written by a single standalone node with exclusive access to the target disk ? or is another node in the cluster trying to do a shadow set merge as the dumpfile is being written to the same disk... ?
In practice, in a SAN it is a good idea to do dump-off-system-disk to a physically local disk to each node. A few years ago I reported a dumpfile write problem to engineering.
If you have a san with dual HBA and dual controllers for each LUN, you have 4 paths to each logical disk. If your dumpfile is on a three member shadow set, you have a total of 12 logical paths to the various members.
The problem is that the path to use is chosen by the console, which can only know about a maximum of 4 paths. It was quite common to find that the console tried all 4 paths it knew about without find a single valid path, so the attempt to write a dump was aborted...
Hence the advice to use a local SCSI disk for a DOSD dumpfile. The disk does not need to be mounted by VMS.


In other words, how long is your particular piece of string... :-)
Robert Brooks_1
Honored Contributor
Solution

Re: Speed of writing dump file and DUMPSTYLE settings

The engineer who works on the crash dump facility has repeatedly said that there is little reason to use a non-compressed or full dump. If one has ever requested a selective, compresssed dump, and all the pertinent information was NOT written, we'd like to hear about it.

Anything that can reduce the amount of data physically written to the disk will certainly speed up the dumping process; the CPU overhead is considered to be negligible (this, again, comes directly from the guy who spends his life optimizing crash dump creation.)


-- Rob (VMS Engineering)
Volker Halle
Honored Contributor

Re: Speed of writing dump file and DUMPSTYLE settings

Cass,


Hence the advice to use a local SCSI disk for a DOSD dumpfile. The disk does not need to be mounted by VMS.


During startup, the DOSD disk should be mounted (in SYCONFIG or SYLOGICALS) with a logical name CLUE$DOSD_DEVICE pointing to it. Otherwise, you will not get the CLUE file and the entry in CLUE$HISTORY.

There was a discussion about dumpfile write speed in ITRC quite some time ago. I would encourage anybody, who gets a dump and can time the dump write operation, to just document to necessary data for reference.

http://forums1.itrc.hp.com/service/forums/questionanswer.do?threadId=861610

Volker.
John Travell
Valued Contributor

Re: Speed of writing dump file and DUMPSTYLE settings

> If one has ever requested a selective, compresssed dump, and all the
> pertinent information was NOT written, we'd like to hear about it.
I HAVE seen a crash where pertinent information was NOT saved, and the event had to be written off as unresolveable.
It was a few years ago, while I was working for the UK suppport centre, so may have been fixed since then. Certainly there will not be any evidence preserved for investigation now.
IIRC, the crash revolved around invalid data found in an XQP cache line. At least, that's what the listing said, but being a selective dump the cache lines were not saved so we could never identify what was wrong with that cache line.
I would expect that this situation has not changed, XQP cache could consume quite a lot of memory, and implicitly a fair chunk of dumpfile space.

Has there been any consideration given to enhancing the compression algorithm ? Modern CPU's could easily do more compressing in the time taken to write the compressed data. Provided that a suitable non-lossy algorithm with a sufficiently higher compression ratio can be found, it would only require a means to identify to the relevant tools which algorithm to use when accessing the file.
Cass Witkowski
Trusted Contributor

Re: Speed of writing dump file and DUMPSTYLE settings

Thanks all for your input.

Regards,

Cass