On 5/2/12 9:28 AM, Christopher Arnold wrote:
On that note, can someone explain the difference between the fragment setting and the min_average size setting? We have had this overflow issue when setting the fragment size to 2M and the min_average to 10M or larger. If I set the fragment size higher than 2M the traffic server performance falls off a cliff or stalls out completely.
The min_average_size basically dictates how many directory entries it will create for the cache disks. Divide your disk sizes by this setting, and you get ballpark the number of files it can hold. If you average object size is 10MB, set min_average size to something less than that, say 1MB. Easiest is to think of this is how you create a FS on Unix, and allocate a max number of I-Nodes.
The segment size is how it organizes large files, they get split up into write segments. (John can probably provide a more intelligent explanation). You can't set this too large, if I recall the discussions, it has to be smaller than a hardcoded AGG_BUFF setting size of 4MB (I made a patch to make this configurable, but I moved it out to "later" since we don't need it yet I don't think).
At the moment I have both set to 2M and have not seen the issue resurface although it's too early to tell. _stats reports that I have a max of 1M direntries with the above config.
Seems reasonble, assuming you have ~2TB of disk cache. (2TB / 2MB ~= 1M). It's important to tune this right, to avoid wasting unecessary memory and disk I/O. Seems like you probably have it set pretty well now (but if you see those warnings, you probably have set it too low).
Cheers, -- Leif
