当前位置:网站首页>CIPU, what impact does it have on the cloud computing industry?
CIPU, what impact does it have on the cloud computing industry?
2022-08-05 06:15:00 【Chengyun Technology】
I once helped a friend test the packet filtering performance of a firewall.In the era when switches were rare and expensive, this firewall claimed to be able to achieve 100M packet filtering performance.However, in order to simulate the bypass monitoring mode during the test, my friend used a 10M HUB to build a network environment.When I discovered this problem, I also explained to my friends why I couldn't use a 10M HUB to test a 100M firewall.I figured out the speed comparison of various types of computer memory: memory > hard disk > network.The problem is that the performance of the buffer overflow is the network + hard disk, which is faster than the network + network when the proxy is used.
I have come to realize that an important factor affecting computing speed is bandwidth.This bandwidth includes network bandwidth, memory bandwidth, memory access bandwidth, and so on.Modern processors are so fast that many times the processor runs out of instructions and has to wait for the next data item to arrive.This is why the computer has been promoting the use of faster internal bandwidth, the memory bandwidth has been upgraded to DDR6, and the network bandwidth has also changed from the 10M network HUB when I first came into contact with the computer to the 1G internal network at my home now.The office has a 10G intranet.
In real computing tasks, the common optimization state is to full bandwidth.For example, for an HTTP server, it is basically impossible to use up the CPU, but as long as the network card can be fully run, it will be fine.In FPGA-type applications, the memory access bandwidth is often used up.
It is such an environment that servers are constantly trying to upgrade various bandwidths.As instructed by friends who have experience in server computing power, the main frequency of the server CPU is not high, even lower than that of the desktop computer.However, the server CPU has a very huge bandwidth, including support for multi-channel memory, massive PCIE channels, support for multiple high-performance network cards, and so on.The source of these optimization requirements is that many computing tasks require very large bandwidth, otherwise the computing power of the CPU can only be starved, and there is not enough input or output data.
In the process of server's continuous pursuit of greater bandwidth, cloud computing platform represents the peak of optimization of high-performance servers, and is also pursuing more extreme bandwidth performance.
Alibaba Cloud's newly launched CIPU is a new peak for high-bandwidth optimization.
CIPU is an infrastructure processor developed by Alibaba Cloud. It replaces the CPU in the data center and becomes the center of the cloud computing architecture. It is a dedicated processor designed for the Alibaba Cloud Feitian cloud operating system.
CIPU manages data center hardware downward and accelerates computing, storage, network and other resources, providing extreme bandwidth capabilities, so that the computing process is no longer limited by bandwidth.
CIPU has done a lot not only in bandwidth optimization, but also in improving computing performance.It is also Alibaba Cloud's self-developed chip, which enables various optimizations of cloud computing to go directly to the hardware, avoiding the delay and loss caused by the use of software in the past, which is a leader in the cloud computing industry..
Only a cloud platform that has both cloud computing server hardware and cloud operating system can make such a dedicated processor.
边栏推荐
- Spark source code - task submission process - 6-sparkContext initialization
- IP packet format (ICMP protocol and ARP protocol)
- spark算子-wholeTextFiles算子
- Configuration of TensorFlow ObjecDetectionAPI under Anaconda3 of win10 system
- IP地址及子网的划分
- I/O performance and reliability
- [Day8] (Super detailed steps) Use LVM to expand capacity
- 【Day5】软硬链接 文件存储,删除,目录管理命令
- 账号与权限管理
- Spark源码-任务提交流程之-6-sparkContext初始化
猜你喜欢
Why can't I add a new hard disk to scan?How to solve?
[Day6] File system permission management, file special permissions, hidden attributes
Getting Started 04 When a task depends on another task, it needs to be executed in sequence
Getting Started Doc 06 Adding files to a stream
I/O性能与可靠性
技术分享杂七杂八技术
Hard Disk Partitioning and Permanent Mounting
Transport layer protocol (TCP 3-way handshake)
OpenCV3.0 is compatible with VS2010 and VS2013
spark source code-RPC communication mechanism
随机推荐
IP数据包格式(ICMP协议与ARP协议)
spark operator-parallelize operator
入职前,没想到他们玩的这么花
ACLs and NATs
Getting Started 05 Using cb() to indicate that the current task is complete
Remembering my first CCF-A conference paper | After six rejections, my paper is finally accepted, yay!
【Machine Learning】1 Univariate Linear Regression
[Pytorch study notes] 8. How to use WeightedRandomSampler (weight sampler) when the training category is unbalanced data
idea 常用快捷键
Getting Started Document 09 Standalone watch
Spark source code-task submission process-6.1-sparkContext initialization-create spark driver side execution environment SparkEnv
NIO工作方式浅析
The spark operator - repartition operator
入门文档07 分阶段输出
【机器学习】1单变量线性回归
Apache配置反向代理
Spark source code - task submission process - 4-container to start executor
磁盘管理与文件系统
spark operator - map vs mapPartitions operator
深度 Zabbix 使用指南——来自惨绿少年