Tengxun Cloud Account Recharge: Deep Hard Core Measurement of Multi-task Concurrent Processing Ability of Computing Server
In the cloud computing market, there is a classic "three century problem":
How to choose universal type, memory type and computing type? Tengxun cloud account top-up
Many executives who have just entered the industry to do architecture or lead a team are often prone to fall into a misunderstanding: "It's all cloud servers anyway, so why don't I buy a general-purpose type with more cores and large memory? Isn't the computing server just a CPU with a high dominant frequency and a high loss? As for the separate classification, is it more confident to sell?"
In order to thoroughly understand the real performance of "computing servers" under multi-task concurrent and high load pressure measurement, our team recently made an almost crazy
Hell-level multi-task concurrent stress measurement
. We found one.
Tengxun Cloud's latest generation of computing servers (16 cores and 32G)
, directly transcoding video, AI reasoning, and complex encryption and computing
At the same time pull full run
.
Today's article, without carrying any official rhetoric, uses the most down-to-earth real-life perspective and first-hand measured data to show you the "horrible waist power" of computing servers in multitasking concurrent processing ".
1. why multitasking concurrent, must look for "computation"?
Before we get down to business, let's talk in plain English:
What is the test of multi-task concurrent processing (Multi-tasking Concurrent Processing) on the bottom of the server?
Many people think that multitasking concurrency is "1 CPU core is not enough, then give 10 cores to run together". That's true, but on an ordinary general-purpose server, when multiple heavy computing tasks erupt at the same time, the system often encounters the following two fatal bottlenecks:
Pseudo multi-core and computing power grab (CPU Churning): The basic CPU frequency of an ordinary server may be only $2.5\text{ GHz }$, and there may be "resource sharing" of virtualized hyper-threading at the bottom layer ". When multiple tasks need to calculate power at the same time, CPU cores frequently switch contexts, resulting in a large amount of computing power wasted on "queuing for seats.
Cache anemia (Cache Starvation): Multi-tasking is most afraid that the L3 cache (level 3 cache) in the CPU is not large enough. If task A's data is just put into the cache and is squeezed out by task B before it is finished, the CPU will have to pull data from memory frequently, resulting in a cliff-like drop in performance.
The computing server (Compute-Optimized Instance) is born to crack this dead game. Its core features are:
The ratio of CPU to memory is deadlocked at $1:2$(e.g. 4-core 8G,16-core 32G), throwing every budget at CPU performance.
Exclusive high-frequency processors usually come standard with high-end chips with a maximum turbo boost of $3.5\text{ GHz}$or more.
Has a huge L3 cache per core, ensuring that multiple tasks are concurrent when the number of their own
The data can stay in the cache nearest to the CPU.
2. Measured Adventure: Three "Electric Tiger" Missions Bombing at the same time
To test its limits, we constructed an extreme
Multi-task concurrent mixed scenario
. If ordinary servers had run like this, the operating system might have gone on strike or crashed directly.
📊Our Measured Environment
Test model: computing server (16 cores 32G, exclusive physical core)
Operating System: CentOS Stream 9
Concurrent task combination: Task A (video group): Use FFmpeg to perform H.265 encoding and transcoding on 4 channels of $4\text{K}$ultra-clear videos at the same time (extremely squeezing the arithmetic logic unit ALU of the CPU). Task B (security group): Run a high-frequency Python script to continuously generate RSA-4096 keys and decrypt large files (to squeeze the CPU's bit and integer computing capabilities). Task C(AI Inference Group): Run a lightweight BERT text classification model for continuous concurrent text sentiment analysis (squeeze CPU matrix multiplication with instruction set extensions such as AVX-512).
3. Concurrent Performance: Data Does Not Lie
When the three tasks in the background at the same time hit enter to start the moment, we keep a close eye on the monitoring board.
1. "Robust curve" at 100% full load"
Tencent Cloud account top-up
Under the attack of three "computing power devouring beasts", the server's 16 CPU cores all soared in less than 2 seconds.
100% full load condition
.
If you used to use an ordinary general-purpose server, when you pull the SSH terminal to enter the command, there will usually be obvious stuck, dropped or even refused to connect. But on compute servers, we try to execute
top
Command and view system logs, terminal feedback actually
Extremely silky, without delay
. This means that the underlying layer retains an extremely strong response channel for kernel scheduling and high-priority tasks (such as system interaction).
2. Core index measured comparison
We let this mixed multitasking run continuously for 30 minutes, and compared it horizontally with a general-purpose example of the same specification (16 cores and 64G):
Test metrics and task performance
Common General Example (16-core 64G)
Computational instance (16 cores 32G)
Performance Gap and somatosensory
FFmpeg 4K frame rate (total)
Average of 42 frames per second
78 frames/second average
Increase by about 85%, transcoding speed nearly doubled
RSA Decryption Throughput
2,100 times/second
3,950 times/second
Higher computational purity, far ahead of integer arithmetic
AI Text Reasoning Delay (P99)
142ms (violent fluctuation)
38ms (extremely stable)
Thanks to AVX-512 instruction set optimization
CPU temperature and frequency under high load
Encounter Temperature
Wall, frequency down to 2.6G
Always stable at 3.4G Turbo Boost
Host heat dissipation and power supply is extremely strong
3. Multi-task non-interference "boundary" experience
In the test, we made a small move: in the 15th minute, we suddenly transcoded the number of video tasks
Double it
(from Route 4 to Route 8).
On a general-purpose server, this burst of computing power can cause the "AI inference delay" next door to soar to hundreds of milliseconds in an instant. On the compute server, however, the latency of AI inference jitters only slightly (from 38ms to 45ms) and then immediately returns to normal.
This reflects the powerful computing server.
Multi-threaded hardware isolation and large cache advantages
. Each core is doing its own dirty work, and the hardware-level assembly line is arranged in an orderly manner, and there is no tragedy of "one person occupying the road and blocking the whole line.
4. Depth Begins: Why Is It Multitasking and Concurrency So Strong?
To remove the surface data, we have to look at the three core secrets of computing server multitasking power from the bottom of the technology:
Secret 1: Hardware-level instruction set (AVX-512 / AMX) blessing
Modern computing servers use CPUs that integrate a large number of "advanced vector extension instruction sets" (such as Intel's AVX-512).
An ordinary server calculates a complex mathematical matrix, which requires several steps to go through the pipeline, while the underlying instruction set of a computational server,
Allow the CPU to calculate a large row of data at the same time, just like cutting leeks.
. When running multiple tasks, this hardware-level "cheating artifact" can quickly close specific tasks and free up computing power for other tasks.
Secret 2: No "moisture" of the physical computing power
Many cheap virtualized VPS or general-purpose low-end instances have CPU cores that are "shared" by multiple users at the bottom (also known as oversold).
And the computing servers of large factories, which usually promise
1:1 physical core binding
. The 16 cores are the real 16 physical computing units that are exclusive to you. When multitasking is concurrent, each task is assigned to a truly exclusive "personal bodyguard", so there will naturally be no serious resource tearing.
Secret 3: Gold Memory Ratio (1:2) Reduces Overhead
Some people ask: "Why is it better to have 32G of memory for 16 cores of computing servers and 64G for general-purpose servers?"
This is where the big factory shrewd. Most of the data in computing services (such as compilation, rendering, and encryption) is rotated in the CPU cache at a high frequency and does not require much memory capacity.
Cut off excess memory capacity in exchange for elite memory with higher frequency and lower latency
. This instead reduces the system overhead of the CPU waiting for large memory to empty data.
5. selection combat: how should your multi-tasking business be in place?
After watching our extreme pressure test, you may have already moved. But calm down, computing servers are good, but by no means everything.
Of it. I 've summarized a pragmatic selection formula for you:
Tencent Cloud account top-up
🚀Without hesitation, please directly lock the scenario of [computing server]:
High concurrency Web backend and API gateway: For example, your backend has a large number of business logic judgment, data verification, and permission encryption (Java / Go / Node.js intensive business).
Audio and video processing and multimedia cleaning: every day run FFmpeg video slicing, transcoding, watermark addition, image compression.
Large-traffic scientific computing and batch runs: for example, the financial statements and actuarial models of thousands of users need to be calculated at high concurrency every night.
Lightweight ML deployment: Not worth expensive GPU, CPU is required for efficient and concurrent online AI prediction and NLP text word segmentation.
🛑Listen to my advice, please detour to choose the scene of [universal or memory]:
Highly concurrent non-relational databases (such as Redis):Redis core looks at memory bandwidth and capacity, and a 16-core 32G computing server is "CPU idle and memory crowded" for it ".
Large single e-commerce database (such as MySQL / Oracle): The database needs huge memory to make Buffer Pool. The compute server's memory is too small, causing frequent disk I/O triggers.
Pure file storage and distribution: it is only used to download large files to clients. CPU is idle every day and should add money to buy public network bandwidth and high throughput cloud disks.
6. Summary
If you compare a general-purpose server to an "all-round handyman" who can do everything but is not good at anything, then
The computing server is an "elite special forces" for high-intensity, high-concurrency and high-difficulty computing"
.
In the face of the triple attack of video transcoding, AI inference and high-intensity encryption, the computing server uses it.
Up to 3.4G stable turbo boost, 1:1 exclusive physical computing power and powerful hardware extension instruction set
, handed over an answer close to full marks. It tells us: in the battlefield of multitasking and concurrent processing, it is often not how much memory you have, but how pure your CPU computing power is!
Tencent Cloud account top-up