Junhua wang ai_next_con
3 likes945 views
The document discusses applications of deep learning for search, speech, vision, and machine reading. It describes a large-scale deep learning service for inference and vector search that is in production globally. The service supports heterogeneous hardware and pluggable frameworks to optimize models for different hardware including CPUs, GPUs, and FPGAs. It aims to distribute models optimally across server fleets while matching requirements to hardware.
![DL Scenarios Original Latency
Latency
Target
Optimized Latency
Latency
reduction
Throughput
improvement
Turing Prototype 2 ~100ms 10ms 9ms >10X > 10X
Turing Prototype 3 ~107ms 10ms 4.1ms >20X > 50X
Deep Query Document
Similarity
10~12ms for [query,
1 doc] x 33 docs
6ms
1.5ms for [query, 1 doc];
<6ms for [query, 33 docs]
>6X > 30X
Malta Click Features
10ms for
[query, 1 passage]
x 150 passages
5ms
<1ms for [query, 1 passage];
<5ms for [query, 150 passages]
>10X > 100X
Ads seq2seq model for
query rewriting
51ms 5ms 4ms >10X > 3X
AGI Encoder V2 ~29ms 10ms 5.4ms 5X 5X
RNet (InfoBot + Bing)
~45ms for 1 [query,
passage]
10ms
4.0ms for 1 [query,
passage];
<8.5ms for 20 [query,
passage]
11X > 100X
Bing query tagging 9~16ms on CNTK 3ms 0.95ms 10X > 10X
WideDeepRight Model
(TP3 L1)
~25ms for [query, 1
title url]
7ms for a
batch size of
33
5.4ms for [query,
33 title url];
10X > 100X
TP3 L2 Classifier 60ms 3ms 3ms 20X 20X
TP3 L1 8ms 3ms 1ms 8X 8X
Optimization](https://image.slidesharecdn.com/junhuawangainextcon-180618175010/85/Junhua-wang-ai_next_con-22-320.jpg)
Junhua wang ai_next_con
- 1. AI NEXTCon |
- 2. Agenda Applications FrameworkOptimization Fundamental Feature Ask Optimize for Application Make optimization Re-usable for others
- 4. Applications (0.78, 0.8, 0.4, 0.3, 0.9,...) (0.75, 0.6, 0.1, 0.7, 0.2,...) … … • Semantic similarity Vector Representation Nearest Neighbor Search in Semantic Space Q: {is it legal for 17 year old to buy a car} Bag of Words Inverted Index Matching car legal … OR AND buy own • L1: BM25F Ranking Posting 1 Posting 2 Posting 3 Posting 4 … buy legal … L2/L3/L4 ReRanking
- 5. Semantic search can help recall issues, nearly a third of relevance DSATs. Query: {how many women voices in Switchboard telephone corpus } Cannot recall the good urls by query term alteration and term match DL model captures full context, and builds semantic meanings into vectors. The query vector and document vector are near in vector space. Applications
- 7. Query: Where's the nearest fruit smoothies Location: Omaha, Nebraska Applications
- 8. Framework Deep Learning Platform DLIS Pluggable Runtime Linux ContainerNative Windows Microsoft CNTK TensorFlowWin DeepCPU TensorFlowLinux Caffe Hardware Accerlation CPU GPUFPGA Self-Serve Portal Model Development toolkit Model Repository Theano ... Workloads Web Text Speech Image Enterprise DLVS Pluggable Runtime HNSW K-D Tree Faiss ANN Index Build on Multi-Tenancy FrontDoor DLIS DLVS • Customizable runtime • Privacy and Compliance Certification • In production globally
- 9. • • • 1Ms QPS, 100s models, 100Bs vectors, 20+ Regions Framework
- 10. Optimal distribution to match model requirements to server fleet Framework Windows Machine SKU-2 Model2 Model2 Linux Machine SKU-4 Windows Machine SKU-1 Model1 Model1 Linux Machine SKU-5 Model6 Windows Machine SKU-3 Model6 Model6 Model6 Model2 Model2 Windows Machine SKU-2 Model2 Model2 Windows Machine SKU-2 Model3 Model4 Model5 Model2 Model2 Model1 Model1 Model1 Model1 Model1 Model1 Model1 Model1 Model1 Model4 Model5 Multiple model instances across multiple machines Multiple model instances share same machine Different Operating System in same bed Different runtime Model7 Linux Machine SKU-5 Model8Model7 Model7 CNTK TensorFlow Windows DeepCPU TensorFlow Linux Different machine SKU in same bed
- 11. Query: coffee in Melbourne Semantic Representation Vector Online Inferencing Batch Inferencing Document 1 Document 2 ... Vector Set Similarity Search Framework
- 12. Vector Recall by Nearest Neighbor Search Search among points in bucket Hash query to this bucket NNG HNSW KD-tree Semantic word 1 Semantic word 2 Semantic word 3 TP-tree & Wang, Jingdong, and Shipeng Li. "Query-driven iterated neighborhood graph search for large scale indexing." Proceedings of the 20th ACM international conference on Multimedia. ACM, 2012. Framework
- 13. Optimization Hardware + Software Acceleration DeepCPU BrainWave / FPGA DeepGPU
- 14. RNN Serving Performance Challenges Language Modeling Machine Translation Machine Reading Comprehension Conversation Bot Speech Recognition … Limited Parallelism Limited Bandwidth • Small batch size • Sequential dependency • Vector-matrix multiplication • Low data reuse 14 Xt-1 Xt Xt+1 Ot-1 Ot Ot+1 St-1 St St+1 W W W U U U V V V Optimization
- 15. 1. Matrix computation: 2. Activation function 3. Operation Fusing 4. Affinity 5. Locality 6. Parallelism 7. Task scheduling Collaborating with Yuxiong He, Minjia Zhang, Samyam Rajbhandari, Wenhan Wang, Microsoft AI and Research. Optimization
- 16. 𝑧𝑡 = 𝜎 𝑊𝑧 𝑥 𝑡 + 𝑈𝑧ℎ 𝑡−1 + 𝑏 𝑧 𝑟𝑡 = 𝜎 𝑊𝑟 𝑥 𝑡 + 𝑈𝑟ℎ 𝑡−1 + 𝑏 𝑟 ℎ 𝑡 = 𝑧𝑡 ∘ ℎ 𝑡−1 + 1 − 𝑧𝑡 ∘ tanh(𝑊ℎ 𝑥 𝑡 + 𝑈ℎ 𝑟𝑡 ∘ ℎ 𝑡−1 + 𝑏ℎ) On a machine with 12 cores… a) 1 core per operation, multiplications done in parallel 1 1 1 1 1 1 time cores 6 12 b) 12 cores per operation, multiplications done sequentially 12 12 12 12 12 12 6 12 cores time many idle cores unbalanced load poor speedup of intra-op parallelism Optimization
- 17. Optimization 𝑧𝑡 = 𝜎 𝑊𝑧 𝑥 𝑡 + 𝑈𝑧ℎ 𝑡−1 + 𝑏 𝑧 𝑟𝑡 = 𝜎 𝑊𝑟 𝑥 𝑡 + 𝑈𝑟ℎ 𝑡−1 + 𝑏 𝑟 ℎ 𝑡 = 𝑧𝑡 ∘ ℎ 𝑡−1 + 1 − 𝑧𝑡 ∘ tanh(𝑊ℎ 𝑥 𝑡 + 𝑈ℎ 𝑟𝑡 ∘ ℎ 𝑡−1 + 𝑏ℎ) On a machine with 12 cores… d) an optimized configuration, reducing latency 6 12 cores time 2 2 3 3 2 6 c) 4 cores per operation 4 4 4 4 4 4 time cores 6 12 1 1 1 2 2 2 1 2 Bad scheduling order ✓ Workload size ✓ Parallelism efficiency ✓ Critical path ✓ Load balancing
- 18. Optimization
- 20. 20 Optimization
- 21. 21 Optimization
- 22. DL Scenarios Original Latency Latency Target Optimized Latency Latency reduction Throughput improvement Turing Prototype 2 ~100ms 10ms 9ms >10X > 10X Turing Prototype 3 ~107ms 10ms 4.1ms >20X > 50X Deep Query Document Similarity 10~12ms for [query, 1 doc] x 33 docs 6ms 1.5ms for [query, 1 doc]; <6ms for [query, 33 docs] >6X > 30X Malta Click Features 10ms for [query, 1 passage] x 150 passages 5ms <1ms for [query, 1 passage]; <5ms for [query, 150 passages] >10X > 100X Ads seq2seq model for query rewriting 51ms 5ms 4ms >10X > 3X AGI Encoder V2 ~29ms 10ms 5.4ms 5X 5X RNet (InfoBot + Bing) ~45ms for 1 [query, passage] 10ms 4.0ms for 1 [query, passage]; <8.5ms for 20 [query, passage] 11X > 100X Bing query tagging 9~16ms on CNTK 3ms 0.95ms 10X > 10X WideDeepRight Model (TP3 L1) ~25ms for [query, 1 title url] 7ms for a batch size of 33 5.4ms for [query, 33 title url]; 10X > 100X TP3 L2 Classifier 60ms 3ms 3ms 20X 20X TP3 L1 8ms 3ms 1ms 8X 8X Optimization
- 24. 24 Original TensorFlow model TensorFlow model with DeepCPU operator Optimization
- 25. F F F L0 L1 F F F L0 Pretrained DNN Model in TF/CNTK/ONNX, etc. Scalable DNN Hardware Microservice BrainWave Soft DPU Instr Decoder & Control Neural FU Network switches FPGAs Optimization
- 26. Optimization
- 27. Production Bing DNN Model 1 CPU only Brainwave accelerated Improvement Model Details GRU 128X200 (X2) + W2Vec LSTM 500X200 (x8) +W2Vec Brainwave accelerated mode is > 10X larger and > 10X lower latencyEnd-to-End latency per Batch 1 request at 95% 9ms 0.85ms Production Bing DNN Model 2 CPU only Brainwave accelerated Improvement Model Details 1D CNN + W2Vec (RNNs removed) 1D CNN + W2Vec + GRU 500x500 (x4) Brainwave accelerated mode is > 10X larger and 3X lower latency End-to-End latency per Batch 1 request at 95% 15ms 5ms Optimization
- 28. Layer GEMM 𝑊𝑖 𝑊𝑓 𝑊𝑜 𝑊𝑐 G*H S 𝑥 𝑡S N G*H N Recurrent GEMM 𝑈𝑖 𝑈𝑓 𝑈 𝑜 𝑈𝑐 H ℎ 𝑡−1H G*H N N G*H S = synthetic_dim H = hidden_dim N = batch_size G = num_gates Optimization
- 29. Optimization RF 𝑊𝑒𝑖𝑔ℎ𝑡𝑠 H G*H H Shared Memory ℎ 𝑡−1 result NN H G*H GRU P4 - FP32, batch_size = 1 *Can add more work in this instance Other variables H N RF Usage SMEM 100 1 3∗100∗100∗4 256∗1024 ≅ 46% 100+3∗100 ∗4 96∗1024 ≅ 2% 20 1 3∗20∗20∗4 256∗1024 ≅ 2% 20+3∗20 ∗4 96∗1024 ≪ 1%
- 30. Summary Significant gain from deep learning in search, speech, vision and machine reading comprehension. Large scale and low latency inference and vector search service in production Heterogenous hardware and pluggable framework support