1. Shredder GPU-Accelerated Incremental Storage and Computation
Pramod Bhatotia§,
Rodrigo Rodrigues§, Akshat Verma¶
§MPI-SWS, Germany
¶IBM Research-India
In this talk, I will present a system called Shredder.
2) Shredder is a gpu-accelerated framework for redundancy elimination.
3) In this work, we have used Shredder as a generic library to accelerate both incremental storage and computation.
Trans) To start with I first motivate, why we need to accelerate redundancy elimination.
USENIX FAST 2012

2. Handling the data deluge
Data stored in data centers is growing at a fast pace
Challenge: How to store and process this data ?
Key technique: Redundancy elimination
Applications of redundancy elimination
Incremental storage: data de-duplication
Incremental computation: selective re-execution
Pramod Bhatotia

3. Redundancy elimination is expensive
Duplicate
File
Chunks
Hash
Yes
Chunking
Hashing
Matching No
Content-based chunking [SOSP’01]
Fingerprint
File
Content
Marker
For large-scale data, chunking easily becomes a bottleneck
Pramod Bhatotia

4. Accelerate chunking using GPUs
GPUs have been successfully applied to compute-intensive tasks
20Gbps (2.5GBps)
Storage
Servers ? 2X
Using GPUs for data-intensive tasks presents new challenges
Pramod Bhatotia

5. Rest of the talk Shredder design Evaluation Case studies Basic design
Background: GPU architecture & programming model
Challenges and optimizations
Evaluation
Case studies
Computation: Incremental MapReduce
Storage: Cloud backup
Pramod Bhatotia

6. Shredder basic design CPU (Host) GPU (Device) Transfer Chunking kernel
Reader
Store
Data for
chunking
Chunked
data
Pramod Bhatotia

7. GPU architecture GPU (Device) Host memory Multi-processor N Device
global
memory
Multi-processor 2
CPU
(Host)
Multi-processor 1
Shared memory
PCI SP SP SP SP
Pramod Bhatotia

8. GPU programming model GPU (Device) Host memory Multi-processor N Input
global
memory
Multi-processor 2
CPU
(Host)
Multi-processor 1
Shared memory
PCI
Output SP SP SP SP
The programming model for GPUs requires designing hybrid application, where the control part runs on the CPUs and the compute-intensive task is offloaded to GPUs.
2) Host co-ordinates the execution of application code on GPU. Where first the host ensures data to operate is available in device memory.
3) Thereafter, the host launches kernel program, where the GPU runtime dynamically schedules the kernel thread to one of the scalar processor.
4) These threads fetches the input data directly from the device memory. Or In case, they want to reuse the data in device memory for multiple operation they can bring the data from global memory to shared memory for fast reuse.
5) Kernel program executes SIMT code, where group of threads share resources on Device (GPU)
Threads
Pramod Bhatotia

9. Scalability challenges
Host-device communication bottlenecks
Device memory conflicts
Host bottlenecks
(See paper for details)
Recall – earlier a basic desing gives us 2x, we can do better by solving 3 challenges.
Pramod Bhatotia

10. Host-device communication bottleneck
Challenge # 1
CPU (Host)
GPU (Device)
Main
memory
Chunking
kernel
Device
global
memory
Transfer
PCI
Reader
I/O
Synchronous data transfer and kernel execution
Cost of data transfer is comparable to kernel execution
For large-scale data it involves many data transfers
Pramod Bhatotia

11. Asynchronous execution
GPU (Device)
CPU (Host)
Asynchronous
copy
Device global memory
Main
memory
Transfer
Buffer 1
Buffer 2
Pros: + Overlaps communication with computation + Generalizes to multi-buffering Cons: - Requires page-pinning of buffers at host side
Copy to
Buffer 1
Copy to
Buffer 2
Time
Compute
Buffer 1
Compute
Buffer 2
Pramod Bhatotia

12. Circular ring pinned memory buffers
CPU (Host)
Memcpy
GPU (Device)
Asynchronous
copy
Device global
memory
Pinned circular
Ring buffers
Pageable
buffers
Pramod Bhatotia

13. Device memory conflicts
Challenge # 2
CPU (Host)
GPU (Device)
Main
memory
Chunking
kernel
Device
global
memory
Transfer
PCI
Reader
I/O
Pramod Bhatotia

14. Accessing device memory
Device global memory
Cycles
Thread-1
Thread-2
Thread-3
Thread-4
Few
cycles
Device shared memory
SP-1
SP-2
SP-3
SP-4
Multi-processor
Pramod Bhatotia

15. Memory bank conflicts
Device global memory
Un-coordinated accesses to global memory lead to a large number of memory bank conflicts
Device shared memory SP SP SP SP
Multi-processor
Pramod Bhatotia

16. Accessing memory banks
Interleaved memory
Bank 0
Bank 1
Bank 2
Bank 3
Memory
address 4 8 1 5 2 6 3 7
Chip
enable OR
MSBs
LSBs
Data out
Address
Pramod Bhatotia

17. Memory coalescing Device global memory Thread # Time Memory coalescing1 2 3 4
Time
Memory
coalescing
Device shared memory
Pramod Bhatotia

18. Processing the data Device shared memory Thread-1 Thread-2 Thread-3
Pramod Bhatotia

19. Outline
Shredder design
Evaluation
Case-studies
Pramod Bhatotia

20. Evaluating Shredder Goal: Determine how Shredder works in practice
How effective are the optimizations?
How does it compare with multicores?
Implementation
Host driver in C++ and GPU in CUDA
GPU: NVidia Tesla C2050 cards
Host machine: Intel Xeon with12 cores
(See paper for details)
Pramod Bhatotia

21. Shredder vs. Multicores5X
Pramod Bhatotia

22. Outline Shredder design Evaluation Case studies
Computation: Incremental MapReduce
Storage: Cloud backup
(See paper for details)
Pramod Bhatotia

23. Incremental MapReduce
Read input
Map tasks
Reduce tasks
Write output
Pramod Bhatotia

24. Unstable input partitions
Read input
Map tasks
Reduce tasks
Write output
Pramod Bhatotia

25. GPU accelerated Inc-HDFS
Input file
copyFromLocal
Shredder
HDFS Client
Content-based chunking
Split-1
Split-2
Split-3
Pramod Bhatotia

26. Related work GPU-accelerated systems Incremental computations
Storage: Gibraltar [ICPP’10], HashGPU [HPDC’10]
SSLShader[NSDI’11], PacketShader[SIGCOMM’10], …
Incremental computations
Incoop[SOCC’11], Nectar[OSDI’10], Percolator[OSDI’10],…
Pramod Bhatotia

27. Conclusions GPU-accelerated framework for redundancy elimination
Exploits massively parallel GPUs in a cost-effective manner
Shredder design incorporates novel optimizations
More data-intensive than previous usage of GPUs
Shredder can be seamlessly integrated with storage systems
To accelerate incremental storage and computation
Pramod Bhatotia

28. Thank You!