[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: Poor Performance WhenNumber of Files > 1M

Ric Wheeler wrote:
Eric Sandeen wrote:
John Kalucki wrote:

Performance seems to always map directly to the number of files in the ext3 filesystem.

After some initial run-fast time, perhaps once dirty pages begin to be written aggressively, for every 5,000 files added, my files created per second tends to drop by about one. So, depending on the variables, say with 6 RAID10 spindles, I might start at ~700 files/sec, quickly drop, then more slowly drop to ~300 files/sec at perhaps 1 million files, then see 299 files/sec for the next 5,000 creations, 298 files/sec, etc. etc.

As you'd expect, there isn't much CPU utilization, other than iowait, and some kjournald activity.

Is this a known limitation of ext3? Is expecting to write to O(10^6)-O(10^7) files in something approaching constant time expecting too much from a filesystem? What, exactly, am I stressing to cause this unbounded performance degradation?

I think this is a linear search through the block groups for the new
inode allocation, which always starts at the parent directory's block
group; and starts over from there each time.  See find_group_other().

So if the parent's group is full and so are the next 1000 block groups,
it will search 1000 groups and find space in the 1001st.  On the next
inode allocation it will re-search(!) those 1000 groups, and again find
space in the 1001st.  And so on.  Until the 1001st is full, and then
it'll search 1001 groups and find space in the 1002nd... etc (If I'm
remembering/reading correctly, but this does jive with what you see.).

I've toyed  with keeping track (in the parent's inode) where the last
successful child allocation happened, and start the search there.  I'm a
bit leery of how this might age, though... plus I'm not sure if it
should be on-disk or just in memory.... But this behavior clearly needs
some help.  I should probably just get it sent out for comment.


I run a very similar test, but normally run with a synchronous write work load (i.e., fsync before close). In my testing, you will see a slow but gradual decline in the files/sec. For example, on a 1TB S-ATA drive, the latest test run started off at a rate of 22 files/sec (each file is 40k) and is currently chugging along at a bit over 17 files/sec when it has hit 2.8 million files in one directory. I am using the ext3 run to get a baseline for a similar run of xfs and btrfs.

One other random tuning thought - you can help by writing into separate directories, but you will need to make sure that you don't produce a random write pattern when you select your target subdirectory. I think that the use case mentioned using a hashed directory structure which is fine, but you want to hash in a way that writes into a shared subdirectory for some period of time (say get a rotation of every X files or Y seconds). Easiest way to do this is to use a GUID with a time stamp and hash on the time stamp bits.

Note that there is a multi-threaded performance bug in ext3 (Josef Bacik had looked at fixing this) which throttles writes/sec down to around 230 when you do synchronous transactions so you might be hitting that as well.


Unfortunately, I don't have the opportunity to limit the directories. My application is taking random-ish data and organizing it into logical groups for subsequent quick reading. But I did take your suggestion into account and it contains what seems to be the important nugget -- too many active directories makes a bad situation worse.

But still, my test reaches a steady state of active directories pretty quickly -- or so I'd like to think. The performance does indeed continue to creep downwards.

I'm doing everything single-threaded. Introducing a second thread seems to be an immediate disaster, even though I'm stripped across 3 disks. Unfortunate. Perhaps moving the journal to another filesystem would allow better multi-threaded throughput, but I'm not sure that this is important to me.

xfs, zfs, btrfs, and reiser could be attractive for my use-case.

Thanks for your response,

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]