Re: readdir() scalability (was Re: [RFC ] dictionary optimizations)

[Xavier Hernandez <xhernandez@xxxxxxxxxx>]

Al 06/09/13 20:43, En/na Anand Avati ha escrit:

On Fri, Sep 6, 2013 at 1:46 AM, Xavier Hernandez <xhernandez@xxxxxxxxxx> wrote:

Al 04/09/13 18:10, En/na Anand Avati ha escrit:

On Wed, Sep 4, 2013 at 6:37 AM, Xavier Hernandez <xhernandez@xxxxxxxxxx> wrote:

Al 04/09/13 14:05, En/na Jeff Darcy ha escrit:

On 09/04/2013 04:27 AM, Xavier Hernandez wrote:

I would also like to note that each node can store multiple elements.
Current implementation creates a node for each byte in the key. In my
implementation I only create a node if there is a prefix coincidence between
2 or more keys. This reduces the number of nodes and the number of
indirections.

Whatever we do, we should try to make sure that the changes are profiled
against real usage.  When I was making my own dict optimizations back in March
of last year, I started by looking at how they're actually used. At that time,
a significant majority of dictionaries contained just one item. That's why I
only implemented a simple mechanism to pre-allocate the first data_pair instead
of doing something more ambitious.  Even then, the difference in actual
performance or CPU usage was barely measurable.  Dict usage has certainly
changed since then, but I think you'd still be hard pressed to find a case
where a single dict contains more than a handful of entries, and approaches
that are optimized for dozens to hundreds might well perform worse than simple
ones (e.g. because of cache aliasing or branch misprediction).

If you're looking for other optimization opportunities that might provide even
bigger "bang for the buck" then I suggest that stack-frame or frame->local
allocations are a good place to start.  Or string copying in places like
loc_copy.  Or the entire fd_ctx/inode_ctx subsystem.  Let me know and I'll come
up with a few more.  To put a bit of a positive spin on things, the GlusterFS
code offers many opportunities for improvement in terms of CPU and memory
efficiency (though it's surprisingly still way better than Ceph in that regard).

Yes. The optimizations on dictionary structures are not a big improvement in the overall performance of GlusterFS. I tried it on a real situation and the benefit was only marginal. However I didn't test new features like an atomic lookup and remove if found (because I would have had to review all the code). I think this kind of functionalities could improve a bit more the results I obtained.

However this is not the only reason to do these changes. While I've been writing code I've found that it's tedious to do some things just because there isn't such functions in dict_t. Some actions require multiple calls, having to check multiple errors and adding complexity and limiting readability of the code. Many of these situations could be solved using functions similar to what I proposed.

On the other side, if dict_t must be truly considered a concurrent structure, there are a lot of race conditions that might appear when doing some operations. It would require a great effort to take care of all these possibilities everywhere. It would be better to pack most of these situations into functions inside the dict_t itself where it is easier to combine some operations.

By the way, I've made some tests with multiple bricks and it seems that there is a clear speed loss on directory listings as the number of bricks increases. Since bricks should be independent and they can work in parallel, I didn't expected such a big performance degradation.

The likely reason is that, even though bricks are parallel for IO, readdir is essentially a sequential operation and DHT has a limitation that a readdir reply batch does not cross server boundaries. So if you have 10 files and 1 server, all 10 entries are returned in one call to the app/libc. If you have 10 files and 10 servers evenly distributed, the app/libc has to perform 10 calls and keeps getting one file at a time. This problem goes away when each server has enough files to fill up a readdir batch. It's only when you have too few files and too many servers that this "dilution" problem shows up. However, this is just a theory and your problem may be something else too..

I didn't know that DHT was doing a sequential brick scan on readdir(p) (my fault). Why is that ? Why it cannot return entries crossing a server boundary ? is it due to a technical reason or is it only due to the current implementation ?

I've made a test using only directories (50 directories with 50 subdirectories each). I started with one brick and I measured the time to do a recursive 'ls'. Then I sequentially added an additional brick, up to 6 (all of them physically independent), and repeated the ls. The time increases linearly as the number of bricks augments. As more bricks were added, the rebalancing time was also growing linearly.

I think this is a big problem for scalability. It can be partially hidden by using some caching or preloading mechanisms, but it will be there and it will hit sooner or later.

Note that Brian Foster's readdir-ahead patch should address this problem to a large extent. When loaded on top of DHT, the prefiller effectively collapses the smaller chunks returned by DHT into a larger chunk requested by the app/libc.

I've seen it, however I think it will only partially mitigate and hide an existing problem. Imagine you have some hundreds or a thousand of bricks. I doubt readdir-ahead or anything else can hide the enormous latency that the sequential DHT scan will generate in that case.

The main problem I see is that the full directory structure is read many times sequentially. I think it would be better to do the readdir(p) calls in parallel and combine them (possibly in background). This way the time to scan the directory structure would be almost constant, independently of the number of bricks.

The design of the directory entries in DHT makes this essentially a sequential operation because entries from servers are appended, not striped. What I mean is, the logical ordering of 

All entries in a directory = All files and dirs in 0th server + All files (no dirs) in 1st server + All files (no dirs) in 2nd server + .. + All files (no dirs) in N'th server.

in a sequential manner. If we read the entries of 2nd server along with entries of 1st server, we cannot "use" it till we finish reading all entries of 1st server and get EOD from it - which is why readdir-ahead is a more natural solution than reading in parallel for the above design.

As I understand it, what the read-ahead translator does is to collect one or more answers from the DHT translator and combine them to return a single answer as big as possible. If that is correct, it will certainly reduce the number of readdir calls from application, however I think it will still have a considerable latency when used on big clusters. Anyway I don't have any measurement or valid argument to support this, so lets see how readdir-ahead works in real environments before discussing about it.

Also, this is a problem only if each server has fewer entries than what can be returned in a single readdir() request by the application. As long as the server has more than this "minimum threshold" of number of files, the number of batched readdir() made by the client is going to be fixed, and those various requests will be spread across various servers (as opposed to, sending them all to the same server).

I've seen customers with large amounts of empty, or almost empty, directories. Don't ask me why, I don't understand it either...

So yes, as you add servers for a given small set of files the scalability drops, but that is only till you create more files, when the # of servers stop mattering again.

Can you share the actual numbers from the tests you ran?

I've made the tests in 6 physical servers (Quad Atom D525 1.8 GHz. These are the only servers I can use regularly to do tests) connected through a dedicated 1 Gbit switch. Bricks are stored in 1TB SATA disks with ZFS. One of the servers was also used as a client to do the tests.

Initially I created a volume with a single brick. I initialized the volume with 50 directories with 50 subdirectories each (a total of 2500 directories). No files.

After each test, I added a new brick and started a rebalance. Once the rebalance was completed, I umounted and stopped the volume and restarted it again.

The test consisted of 4 'time ls -lR /<testdir> | wc -l'. The first result was discarded. The result shown below is the mean of the other 3 results.

1 brick: 11.8 seconds
2 bricks: 19.0 seconds
3 bricks: 23.8 seconds
4 bricks: 29.8 seconds
5 bricks: 34.6 seconds
6 bricks: 41.0 seconds
12 bricks (2 bricks on each server): 78.5 seconds

The rebalancing time also grew considerably (these times are the result of a single rebalance. They might not be very accurate):

From 1 to 2 bricks: 91 seconds
From 2 to 3 bricks: 102 seconds
From 3 to 4 bricks: 119 seconds
From 4 to 5 bricks: 138 seconds
From 5 to 6 bricks: 151 seconds
From 6 to 12 bricks: 259 seconds

The number of disk IOPS didn't exceed 40 in any server in any case. The network bandwidth didn't go beyond 6 Mbits/s between any pair of servers and none of them reached 100% core usage.

Xavi

Avati

_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxx
https://lists.nongnu.org/mailman/listinfo/gluster-devel