Fragmentation in Large Object Repositories Russell Sears Catharine
Fragmentation in Large Object Repositories Russell Sears Catharine van Ingen CIDR 2007 This work was performed at Microsoft Research San Francisco with input from the NTFS and SQL Server teams
Background • Web services store large objects for users – eg: Wikipedia, Flickr, You. Tube, GFS, Hotmail • Replicate BLOBs or files Clients Application Servers – No update-in-place • Benchmark before deployment – Then, encounter storage performance problems • We set out to make some sense of this DB (metadata) Object. Stores Object Stores Object Stores Replication / Data scrubbing
Problems with partial updates • Multiple changes per application request – Atomicity (distributed transactions) • Most updates change object size – Must fragment, or relocate data • Reading / writing the entire object addresses these issues
Experimental Setup • Single storage node • Compared filesystem, database – NTFS on Windows Server 2003 R 2 – SQL Server 2005 beta • Repeatedly update (free, reallocate) objects – Randomly chose sizes, objects to update – Unrealistic, easy to understand • Measured throughput, fragmentation
Reasoning about time • Existing metrics – Wall clock time: Requires trace to be meaningful, cannot compare different workloads – Updates per volume: Coupled to volume size Storage Age: Average number of updates per object
Read performance 12 – SQL good small object performance (inexpensive opens) – NTFS significantly faster with objects >>1 MB • SQL degraded quickly • NTFS small object performance was low, but constant Read Throughput (MB/s) • Clean system Updates per object 0 2 4 10 8 6 4 2 0 NTFS SQL 256 KB Objects NTFS SQL 1 MB Objects
10 MB object fragmentation 40 • NTFS approaching asymptote • SQL Server degrades linearly SQL Server 35 NTFS Fragments/object 30 25 20 15 – No BLOB defragmenter 10 5 0 0 1 2 3 4 5 6 Storage Age 7 8 9 10
Rules of Thumb • Classic pitfalls – Low free space (< 10%) – Repeated allocation and deallocation (High storage age) • One new problem – Small volumes (< 100 -1000 x object size) • Implicit tuning knobs – Size of write requests
Append is expensive! • Neither system can take advantage of final object size during allocation • Both API’s provide “append” – Leave gaps for future appends – Place objects without knowing length • Observe same behavior with single and random object sizes
Conclusions • Get/put storage is important in practice • Storage – Metric for comparing implementations and workloads – Fragmentation behaviors vary significantly • Append leads to poor layout
----BACKUP SLIDES----
Theory vs. Practice • Theory focuses on contiguous layout of objects of known size • Objects that are allocated in groups are freed in groups – Good allocation algorithms exploit this – Generally ignored for average case results – Leads to pathological behavior in some cases
Small volumes • Small objects / Large volumes – Percent free space • Large objects / Small volumes – Number of free objects
Efficient Get/Put • No update-in-place – Partial updates complicate apps – Objects change size • Pipeline requests – Small write buffers, I/O Parallelism Application server 42 3 1
Lessons learned • Target systems avoid update-in-place No use for database data models • Quantified fragmentation behavior – Across implementations, workloads • Common API’s complicate allocation – Filesystem / BLOB API is too expressive
Application server 3421
Example systems • Share. Point – Everything in the database, one copy per version • Wikipedia – One blob per document version; images are files • Flickr / You. Tube • GFS – Scalable append; chunk data into 64 MB files • Hotmail – Each mailbox is stored as a single opaque BLOB
The folklore is accurate, so why do application designers… …benchmark, then deploy the “wrong” technology? …switch to the “right one” a year later? …then switch back? !? Performance problems crop up over time
Conclusions • Existing systems vary widely – Measuring clean systems is inadequate, but standard practice • Support for append is expensive • Unpredictable storage is difficult to reliably scale and manage – See paper for more information about predicting and managing fragmentation in existing systems
Comparing data layout strategies • Study the impact of – Volume size – Object size – Workload – Update strategies – Maintenance tasks – System implementation • Need a metric that is independent of these factors
Related work • Theoretical results – Worst case performance is unacceptable – Average case good for certain workloads – Structure in deallocation requests leads to poor real-world performance • Buddy system – Place structural limitations on file layout – Bounds fragmentation, fails on large files
Introduction • Content-rich web services require large, predictable and reliable storage • Characterizing fragmentation behavior • Opportunities for improvement
Data intensive web applications • Simple data model (BLOBs) – Hotmail: user mailbox – Flickr: photograph(s) Clients Application Servers • Replication – Instead of backup – Load balancing – Scalability DB (metadata) Object. Stores Object Stores Object Stores Replication / Data scrubbing
Databases vs. Filesystems • Manageability should be primary concern – No need for advanced storage features – Disk bound • Folklore – File opens are slow – Database interfaces stream data poorly
Clean system performance 12 Read throughput (MB/sec) SQL Server 10 • Single node NTFS – Used network API’s 8 6 • Random workload 4 – Get/put one object at a time 2 0 256 K 512 K Object Size 1 M • Large objects lead to sequential I/O
Revisiting Fragmentation • Data intensive web services – Long term predictability – Simple data model: get/put opaque objects • Performance of existing systems • Opportunities for improvement
Introduction • Large object updates and web services – Replication for scalability, reliability – Get / put vs. partial updates • Storage – Characterizing fragmentation behavior – Comparing multiple approaches • State-of-the-art approach: – Lay out data without knowing final object size – Change the interface?
- Slides: 28