Hey

At my company we run a fairly large(by our standard) website whit about 3000 concurrent users, under normal circumstances this runs fine, there is between 50 and 80 attachments and web requests are served in under 100 on average. H

The traffic is "bursty" and at certain intervals eg. when class begins, we receive a lot requestes. These burst result in 100-200 attachments, but no significant change i web request duration.

However every week or two we experience a slowdown where attachment count rises to between 600 and 1400 and web request takes over 20 seconds on average. 

I found the FileSystemCacheSize(Windows only) that instruct firebird/Windows to only use a certain procent of memory for file cache, see CORE-3791. This made me think if something similar was happening on our Linux server. My thought, is that when these bursts, happens and 10, 20, 30 Gb of ram is consumed by the Super Classic process, that the OS evicts pages in a sub optimal way, and perhaps sorting is performed on disk. Once this happens its a negative spiral where queries get slower, and hence more attachments are created.  

I would be very happy to hear from others that might have experienced similar issues, or if you have an idea.

//Thomas Kragh 

Information:

Super Classic 2.5.7 on CentOs

16 cores, 128 Gb

Database size 82 Gb

Page size 16K 

Config(Different from default):

LockMemSize = 5048576

LockHashSlots = 30011

TempCacheLimit = 4294967296

TempBlockSize = 2048576

DefaultDbCachePages = 1024

Firebird is configured to allow 64K open files.