MongoDB tuning on AWS
1 like1,786 views
Ryuji Tamagawa gave a presentation on using MongoDB. He introduced MongoDB as a hybrid of NoSQL and SQL that scales easily. To improve performance on AWS, one must understand MongoDB's memory management, the application's workload pattern, and the size of "hot" data that is frequently accessed. MongoDB does not manage memory itself, instead using memory mapping to leverage the OS buffer. It is important to give the OS enough memory to store hot data and account for indexes. Replication can be used to distribute data across multiple instances for high availability. Performance tests showed SSD-backed instance types had faster write times than non-SSD types for storing Wikipedia data in MongoDB.
MongoDB tuning on AWS
- 2. Information •Tweet: Hashtag #jawsdays #ijaws •Please register you on ijaws on Doorkeeper (Next meetup on Mid April) •There’s the JOB board behind the wall
- 3. Self-Introduction •Ryuji Tamagawa@facebook •tamagawa_ryuji@twitter •Software Developer working in Osaka •Translator (for O’Reilly) •Loves performance tuning
- 4. Introducing MongoDB •Hybrid of NoSQL and RDB •Easily Scales (up to certain point) •Stores JSON document as ‘BSON’ •Has Seconday Index ( on any part of JSON Doc), Query Optimizer •Replication, Sharding ready
- 5. To make MongoDB runs fast on AWS •You have to understand: •its architectural feature of memory management •Workload pattern of your application •Size of your ‘HOT’ data
- 6. What’s the ‘HOT’ data? •‘Hot’ Data is what accessed frequently •Ex: If you simply write data like access logs and transfer them to somewhere else, ‘hot’ spot could be very small •If the collection has indexes, one write can make many places hot
- 7. MongoDB does not manage memory •Most DBMS has built-in MMS, but MongoDB doesn’t. •MongoDB accesses database files through ‘Memory mapped files’: Let the OS manage the buffer Traditional RDB Memory Buffer DB Files MongoDB Memory Mapped DB Files OS App
- 8. The Rules of Thumb about Memory •Give enough memory to the OS to hold ‘HOT’ data •Don’t forget about the indexes •Use dedicated EC2 instances
- 9. Keep your data safe with Replication •Using ReplicaSet, you can distribute your data to many places easily •You have choices to keep your data safe from crashes •EBS or Instance Store : trade off between cost, safety, performance Primary Secondary Secondary Try MongoDB’s Replicaset with: https://bitbucket.org/tamagawa_ryuji/mongodb_replicaset_playground_on_vagrant
- 10. Storage Performance Evaluated • Converted Wikipadia-ja’s page data (about 1,700,000 documents) to JSON • Write them to MongoDB on EC2 from another instance • Data writer is a simple python application with pymongo driver running 4 processes
- 11. Storage Performance Evaluated Instance Type Instance Cost(Spot) Storage Time to finish ebs-normal 0:10:55 ephemeral0 0:07:36 PIOPS 1500 0:08:26 ephemeral0 0:10:22 PIOPS 1500 0:09:02 ephemeral0 0:05:19 m3.large $0.09 m3.xlarge (SSD instance store) $0.16 hi1.4xlarge (Storage Optimized) $0.50
- 12. Comparing Instance Types Instance Type CPU ECU Memory Storage Cost Memory ($/GB) CPU ($/ECU) Storage ($/100GB) m3.medium 1 3 3.75 1 x 4 SSD $0.17 $0.05 $0.06 $4.28 m3.large 2 6.5 7.5 1 x 32 SSD $0.34 $0.05 $0.05 $1.07 m3.xlarge 4 13 15 2 x 40 SSD $0.68 $0.05 $0.05 $0.86 m3.2xlarge 8 26 30 2 x 80 SSD $1.37 $0.05 $0.05 $0.86 m2.xlarge 2 6.5 17.1 1 x 420 $0.51 $0.03 $0.08 $0.12 m2.2xlarge 4 13 34.2 1 x 850 $1.01 $0.03 $0.08 $0.12 m2.4xlarge 8 26 68.4 2 x 840 $2.02 $0.03 $0.08 $0.12 cr1.8xlarge 32 88 244 2 x 120 SSD $4.31 $0.02 $0.05 $1.80 i2.xlarge 4 14 30.5 1 x 800 SSD $1.05 $0.03 $0.08 $0.13 i2.2xlarge 8 27 61 2 x 800 SSD $2.10 $0.03 $0.08 $0.13 i2.4xlarge 16 53 122 4 x 800 SSD $4.20 $0.03 $0.08 $0.13 i2.8xlarge 32 104 244 8 x 800 SSD $8.40 $0.03 $0.08 $0.13 hs1.8xlarge 16 35 117 24 x 2048 $5.67 $0.05 $0.16 $0.01
- 13. THANK YOU ! YOUR CONTACTS ARE WELCOME !!