of cluster nodes ？ Dynamic-Scheduler Node1 Node2 Kube-scheduler Pod Request Load Level Request Load Level Real Load Level Real Load Level Assigned to Node2 The native K8S scheduling is based on Dynamic-Scheduler Node1 Node2 Kube-scheduler Pod Request Load Level Request Load Level Real Load Level Real Load Level Assigned to Node1 Dynamic-scheduler Node1 has a lower load Dynamic-Scheduler Node1 Node2 Dynamic-scheduler- node-annotator 5m Load Prometheus 1h Load 1d Load 5m Load 1h Load 1d Load telegraf Record to node annotation telegraf Dynamic-Scheduler Predicate Node2 Node3 Node1 5m Load

platform ➔ On the server side, it is a network, consisting of some/many nodes ➔ Every node has a blockchain node in it What An Open Data App Feels Like? “EightFish is a development framework (maybe Protocol node 1 node 2 node 3 node 4 node … Protocol 1 node 1 node 2 node 3 node 4 node … Protocol 2 One Protocol Supports Multiple Applications node 1 node 2 node 3 node 4 node

template struct Stack { struct Node { T t; shared_ptr<Node> next; }; atomicNode>> head; void push_front(T t) { auto p = make_shared<Node>(std::move(t), head.load()); different atomic! struct Node { T t; shared_ptr<Node> next; }; atomicNode>> head; void push_front(T t) { auto p = make_shared<Node>(std::move(t), head.load());

Smart Pointers template class concurrent_stack { struct Node { T t; shared_ptr<Node> next; }; atomic_shared_ptr<Node> head; // in C++11: remove "atomic_" and use special // functions functions every time you touch head public: class reference { shared_ptr<Node> p; }; auto find(T t) const { auto p = head.load(); // C++11: atomic_load(&head) while (p && p->t front() const { return reference(head); } void push_front(T t) { auto p = make_shared<Node>(); p->t = t; p->next = head; while (!head.compare_exchange_weak(p->next, p)){ } //

Storage ） http://www.netlib.org/linalg/html_templates/node91.html 按行压缩（ Compressed Row Storage ） http://www.netlib.org/linalg/html_templates/node91.html 第 1 章：稀疏网格 稠密网格计算粒子经过的格点数量 改用更小的 char 存储 只用一个 有了无边界的稀疏网格，再也不用担心二维数组要分配多大了。 坐标可以无限延伸，甚至可以是负数！比如 (-1,2) 等…… 他会自动在写入时分配 16x16 的子网格，称之为叶节点 (leaf node) ，而这里的 unordered_map 就是充当根节点 (root node) 。 图片解释稀疏的好处 传统稠密二维数组 无边界稀疏分块哈希表 此外，还是按需分配内存，即使被写入的部分奇形怪状也不会浪费内存。 这些被写入的部分被称为激活元素

com/Altinity/clickhouse-operator/master/manifests/oper ator/clickhouse-operator-install.yaml Single-node “cluster” apiVersion: "clickhouse.altinity.com/v1" kind: "ClickHouseInstallation" metadata: name: имеющийся ZK cluster Install Zookeeper: manifests/zookeeper/quick-start-persistent-volume/zookeeper-1-node-create.sh Replication apiVersion: "clickhouse.altinity.com/v1" kind: "ClickHouseInstallation" metadata:

still available at JavaScript layer for platform-specific scenarios o NativeScript modules follow Node module conventions (CommonJS) Example: NativeScript file module BUT HOW DO I TURN THIS INTO

block_size>>>(N, x, y); // … } Ok, about the kernel parameters 10 |Memory“In a typical PC or cluster node today, the memories of the CPU and GPU are physically distinct and separated by the PCI-Express

双重执行模式 • 单机和分布式两套计算系统，在不同的使用 环境中都能达到最佳性能  针对常用算法逐个设计优化方案 • 对于常用算法，跳过固定的编程模型，分 别设计最佳的计算方案 • 例如我们自研的 node2vec 采样算法比现 有技术快了 1 个数量级 海致图计算平台特点 AtlasML 极致的性能 • 支持 CPU/GPU 等异构设备训练 • 特殊设计的高性能图算子库 丰富的算法库