Question Reduce checkpoint duration

[cj_brandt]

I know you don't want to upgrade service packs - but improvements in SP06 for handling traffic from remote clients and the lruskips parameter would help.

Checkpoint duration is 30 seconds with -B2 in use and -B2 is only using 20mb. When -B2 is removed, the checkpoint duration is 10 seconds but read performance drops significantly. That isn't something I have seen before, maybe someone else has. I haven't seen a checkpoint last 20+ seconds when the disk sync took less than a second before either though.

It appears the server is fairly busy between 1 and 1:30. The promon R&D 3 1 - performance indicators screen as well as what Rob mentioned - R&D 2 5 screen would be interesting during that time. Also promon screen 5. I would setup windows perfmon to track disk stats as well as CPU, memory and network. If all the changes to the db are coming from remote clients, there must be a lot of network traffic coming in and going out of the server.

 

[Rob Fitzpatrick]

Checkpoint duration is 30 seconds with -B2 in use and -B2 is only using 20mb.

If I read the original post correctly, the -B is 133.5 GB (35 million blocks) and -B2 is 38 GB (10 million blocks), although I don't know how much is allocated to it.

 

[cj_brandt]

Oh... I read the following "overall buffer pool is now 20m less with -B2 set to zero." and thought 20mb instead of 20 million blocks.

 

[Rob Fitzpatrick]

Oh... I read the following "overall buffer pool is now 20m less with -B2 set to zero." and thought 20mb instead of 20 million blocks.

I missed that statement when I first read that post. Either way though (20 million or 20 MB) I don't understand it. It doesn't jibe with the original post.

 

[TomBascom]

... the rest of the server memory is split between other databases on the same box with about 96GB left for the OS ...

So how many other databases are there? And how are they configured? Could they be contributing to the problem?

Also -- I'm sorry that this is vague, but I know that there was a bug of some sort with Windows and very large amounts of shared memory. I don't recall exactly when it was fixed, nor do I have a more useful description but 10.2B06 had a *lot* of performance stuff in it so it may have been fixed there. If I recall correctly it should have been about that time frame -- I remember talking about it last summer.

 

[FocusIT]

This seems to have been resolved in our test environment by upgrading from SP05 to SP07. PSC also did some testing on their side between the Service Packs and found a direct correlation between checkpoint duration and the size of -B/-B2. They have put it down to poor memory management on Windows Server 2008 which has been worked around in SP07. I personally don't care if it is a Windows Server issue or a bug fix in SP07 as long as it resolves the problem.

I will post again when SP07 has been released to our live server.

Thanks for help.