Rapid, significant slowdown using WiredTiger

[Nick Judson]

Hi All,

I'm prototyping MongoDB 3.0 rc8 via the C# driver (2.0.0-beta2). I have the following workflow:

In my current test I have 50 of the above workflows running concurrently. I'm pushing documents into mongo as fast as it'll take it, then subsequently pulling them out for processing and storing the results. It is very 'queue-like'.

I'm running my tests on a i7-5960X with fast SSDs and 16GB of RAM.

When using the MMap engine, Mongo eats all the memory but settles into a consistent (respectable) speed. Everything seems to run smoothly to the 60+M document area (when I stop my test). I originally was using cursors to read all my data without ever closing them (just calling getNext) and that caused quite a few problems. I noticed some notes about needing to close cursors so I dispose of the read cursors every 5k docs.

Anyways, switching to wiredTiger shows an excellent initial speed and resource use. Very quickly the speed drops off to about 10% of the initial speed. This speed drop appears to happen before memory is exhausted, and restarting Mongo (leaving the wiredtiger files) reverts back to the fast speed (so it doesn't appear to be file related). I've used mongoStat to attempt to find some kind of resource limit, but I can't see any.

My Config File:

storage:

    dbPath: "e:\\mongo"

    journal:

        enabled: true

    engine: "wiredTiger"

    wiredTiger:

        engineConfig: 

            cacheSizeGB: 10

            journalCompressor: snappy

        collectionConfig: 

            blockCompressor: snappy 

systemLog:

    destination: file

    path: "c:\\mongo\\mongodb.log"

    logAppend: false

    timeStampFormat: iso8601-utc

net:

    bindIp: nickwork-pc

    port: 27017

I've attached my log file and you can see that the query times appear to linearly increase as time goes on. The SSD that the data files are stored on is a very fast PCI-4X samsung drive. 

I've also attached a MongoStat log and you can see that speeds start at 15k/15k insert/update and degrade to 2k/2k within about 4 minutes.

Now this may very well be something I'm doing wrong on my side since I'm newish to Mongo, however since I don't see this with MMap and it appears the query times do increase. I've confirmed the same behavior on other PCs, some with spindles.

Any guidance would be appreciated!

PS - I also have MMS for this if it would be useful.

[Asya Kamsky]

If anyone else is interested, we are tracking this in https://jira.mongodb.org/browse/SERVER-17299

Asya