Optimal Canary Deployments using Istio and how it scores over Spring Cloud and Kubernetes Presented by Archna Gupta What is a Canary Release or Deployment? • A canary deployment, or canary release

· , k hj(ω) = 0, j = 1, · · · , l with variable ω ∈ Rn, domain D = �k i=1 domgi ∩�l j=1 domhj, optimal value p∗ Objective function f (ω) k inequality constraints gi(ω) ≤ 0, i = 1, · · · , k l equality December 28, 2021 19 / 82 The Lower Bounds Property If α ⪰ 0, then G(α, β ) ≤ p∗, where p∗ is the optimal value of the primal problem Proof: If ˜ω is feasible and α ⪰ 0, then f (˜ω) ≥ L(˜ω, α, β ) ≥ inf Find the best low bound on p∗, obtained from Lagrange dual function A convex optimization problem (optimal value denoted by d∗) α, β are dual feasible if α ⪰ 0, (α, β ) ∈ dom G and G > −∞ Often simplified

for some α and β Theorem 1. Lower Bounds Property: If α ⪰ 0, then G(α, β ) ≤ p∗ where p∗ is the optimal value of the (original) primal problem defined by (9)∼(11). Proof. If �ω is feasible, then we have (9)∼(11) as follows maxα,β G(α, β ) (14) s.t. α ⪰ 0, ∀i = 1, · · · , k (15) We denote by d∗ the optimal value of the above Lagrange dual problem. The weak duality d∗ ≤ p∗ always holds for all optimization optimizing its dual problem. 2.2.2 Complementary Slackness Let ω∗ be a primal optimal point and (α∗, β ∗) be a dual optimal point. Theorem 2. Complementary Slackness: If strong duality holds, then α∗

training by adding only three lines of Python to an existing FP32 (default) script. AMP will select an optimal set of operations to cast to FP16. FP16 operations require 2X reduced memory bandwidth (resulting training by adding only three lines of Python to an existing FP32 (default) script. AMP will select an optimal set of operations to cast to FP16. FP16 operations require 2X reduced memory bandwidth (resulting training by adding only three lines of Python to an existing FP32 (default) script. AMP will select an optimal set of operations to cast to FP16. FP16 operations require 2X reduced memory bandwidth (resulting

elements and a fixed memory budget, how many hash functions do we need in order to minimize Pfp? Optimal number of hash functions 1 0 1 1 1 0 0 1 1 1 1 0 1 1 n bits h1 h2 hk … k hash functions 1 0 elements and a fixed memory budget, how many hash functions do we need in order to minimize Pfp? Optimal number of hash functions After m elements have been inserted to the filter, what is the probability elements and a fixed memory budget, how many hash functions do we need in order to minimize Pfp? Optimal number of hash functions After m elements have been inserted to the filter, what is the probability

that the algorithm needs to process. Unlike traditional algorithm problems where we expect exact optimal answers, machine learning applications can often tolerate approximate responses, since often there higher quality models are deeper, hence will have a higher inference latency. Figure 1-4: Pareto Optimal Models & Pareto Frontier. The green dots are models that have the best tradeoffs for the given objectives which gets a better accuracy while keeping the latency the same (and vice versa). They are pareto-optimal models and together make the pareto-frontier. However, certain models might offer better trade-offs

SSD pct diff Config #1 – MySQL baseline 7,366.55 24,440.57 232% Config #2 – sub‐optimal config 4,478.28 37,802.13 744% pct diff ‐39% 55% 1,000‐wh; 250‐connections SATA NVMe pct diff Config #1 – MySQL baseline 6,668.33 23,175.19 248% Config #2‐ sub‐optimal config 4,457.47 33,857.05 660% pct diff ‐33% 46% Configuration #2 (sub‐optimal, see appendix A) has 3GB of buffer pool, and therefore very little buffer space to be spared as disk cache. In Configuration #3 (MySQL Optimal, see appendix A) we increased

Don’t fret! Let’s start with how to compute the saliency scores. Saliency Scores In the paper Optimal Brain Damage1, LeCun et al. suggested that as much as 50% of the connections (weights) from a large we have a project that relies on it for sparsifying a deep learning model. The authors of the Optimal Brain Damage (OBD) paper approximate the saliency score using a second-derivative of the weights performance." arXiv preprint arXiv:1907.04840 (2019). 1 LeCun, Yann, John Denker, and Sara Solla. "Optimal brain damage." Advances in neural information processing systems 2 (1989). As you can deduce, the

Hyperparameter Optimization (HPO) is the process of choosing values for hyperparameters that lead to an optimal model. HPO performs trials with different sets of hyperparameters using the model as a blackbox. larger number of subspaces or subranges than unimportant parameters that need to be searched for an optimal value. For example, in the US presidential elections, the swing states are the important parameters resources become available or reduced in resource constrained situations. The likelihood of finding the optimal increases with the number of trials. In contrast, the Grid Search has a fixed number of maximum

19 ??? Vasiliki Kalavri | Boston University 2020 20 Optimal parallelism per operator ??? Vasiliki Kalavri | Boston University 2020 20 Optimal parallelism per operator captures upstream operators 2020 20 Optimal parallelism per operator captures upstream operators Aggregated true output rate of operator oj , when oj itself and all upstream ops are deployed with optimal parallelism 2020 20 Optimal parallelism per operator captures upstream operators Aggregated true output rate of operator oj , when oj itself and all upstream ops are deployed with optimal parallelism