TOFFEEプロジェクト
ホームドキュメンテーション更新ビデオ研究ダウンロードスポンサー接触


RESEARCH 》 Live demo - Data Transfer - High bandwidth to Low bandwidth

I always wanted to do some real experiments and research on packet flow patterns from High-bandwidth to Low-bandwidth networks via networking devices. This is something can be analyzed via capturing Network stack buffer data and other parameters, bench-marking, and so on. But eventually the data-transfer nature and other aspects is often contaminated due to the underlying OS and the way Network stack is implemented. So to understand the nature of packet flow from Higher to Lower bandwidth and vice-versa such as Lower to higher bandwidth, I thought I experiment with various tools and things which physically we can observe this phenomena. What we observe in a software test results is not accurate, Operating system is bound by hardware limitations, bound by non-real-time digitized instruction processing and other latencies. And if we cannot understand this fundamental concept in low-speed networks, then we cannot ever understand its complication in a gigabit or high-speed network. Since the high-speed network processing need even more better hardware and OS capabilities. In this case let us assume everything we are trying to experiment is in OS layer (i.e) not in a dedicated hardware or a complete hardware packet processing platform.
network_research_live_demo_kiran

At times there are cases when packet processing happens in Linux Kernel networking subsystem, at each phase there can be a packet buffer or a packet queue. This "phase" can be a module or a sub-module or a component or individual tiny network stacks and so on. So whenever such a packet queue exists we can assume it almost represents a funnel. Since a funnel will have a interface or guide or hole to pour water/liquid. This is similar to packets getting added to the packet queue. And the funnel has the bottom hole or interface where the liquid pours out (exits) of funnel. This again represents a packet queue where the stored packets are taken out (or fetched) and sent out for further processing. The amount of time liquid spends inside the funnel or the size of funnel more or less represents the packet queue length. It is quite common that we can change these parameters in Linux Kernel via /proc interface.

Suppose if you do the ifconfig command you can find the default txqueuelen parameter.

kiran@desktop-i7-5820k:~$ ifconfig
enx00808e8e90f4 Link encap:Ethernet  HWaddr 00:80:8e:8e:90:f4  
          UP BROADCAST MULTICAST  MTU:1500  Metric:1
          RX packets:0 errors:0 dropped:0 overruns:0 frame:0
          TX packets:0 errors:0 dropped:0 overruns:0 carrier:0
          collisions:0 txqueuelen:1000 
          RX bytes:0 (0.0 B)  TX bytes:0 (0.0 B)

eth0      Link encap:Ethernet  HWaddr fc:aa:14:98:cb:66  
          inet addr:192.168.0.101  Bcast:192.168.0.255  Mask:255.255.255.0
          inet6 addr: fe80::feaa:14ff:fe98:cb66/64 Scope:Link
          UP BROADCAST RUNNING MULTICAST  MTU:1500  Metric:1
          RX packets:156020 errors:0 dropped:0 overruns:0 frame:0
          TX packets:104413 errors:0 dropped:0 overruns:0 carrier:0
          collisions:0 txqueuelen:1000 
          RX bytes:165547621 (165.5 MB)  TX bytes:15061686 (15.0 MB)
          Interrupt:20 Memory:fb100000-fb120000 

lo        Link encap:Local Loopback  
          inet addr:127.0.0.1  Mask:255.0.0.0
          inet6 addr: ::1/128 Scope:Host
          UP LOOPBACK RUNNING  MTU:65536  Metric:1
          RX packets:65256 errors:0 dropped:0 overruns:0 frame:0
          TX packets:65256 errors:0 dropped:0 overruns:0 carrier:0
          collisions:0 txqueuelen:0 
          RX bytes:8107654 (8.1 MB)  TX bytes:8107654 (8.1 MB)

In this case the "eth0" interface is my inbuilt motherboard Gigabit Ethernet NIC card. "enx00808e8e90f4" is my USB2 to 100Mbps Ethernet NIC card adapter. And the "lo" is the local host. You can notice the txqueuelen is 1000. txqueuelen is nothing but TX (transmit) queue length. Although as end-users you may see only this in the outside world. But within the Linux kernel you may find packet queues at intermediate levels too. And this almost look like cascading funnels. And each time the funnel sizes varies. This is the reason I took this screenshot of my research video where I am trying to show how they cascade.
network_research_live_demo_kiran_2

Now lets assume you hold a funnel (of average size) as shown and pour water from a tea cup. The capacity of the tea cup is almost the same as funnel's capacity. No matter how fast you pour water, the water never overflows from the top of the funnel brim. This is almost the same case as producer-consumer situation. Where producer-consumer is at the same rate. This is optimal and a good sign.
Note: to highlight the water in the video, I have added red fountain pen ink in the water. So that it is clearly visible.
network_research_live_demo_kiran_3

Now let us imagine we pour water with a large jar as shown in the screenshot. The funnel size is the same, but the water is now being added at a higher rate than what funnel can handle. In this case after a certain point the water will start to overflow from the funnel's top brim. This represents the lack of processing power in your networking device, than the network bandwidth or in other words you are receiving packets at higher rate but you lack the processing power, and at the same time your packet buffers are NOT configured optimally. And the overflow of water from the top of the funnel's brim represents packet drops.

Remember always packet drops happens in a device:

  • anytime if there is congestion
  • if there is data flow from high bandwidth to low-bandwidth
  • some processing delays
  • inadequate buffer capacity (buffer overflow)
  • and so on
network_research_live_demo_kiran_4

In the same way now lets experiment with a small syringe such as Insulin syringe. Insulin syringes have extremely fine needle. And this is a perfect way to obstruct (or add resistance) to liquid flow. As shown in the screenshot each time you such the ink inside the syringe, once you pull the piston, the ink will continue to flow into the syringe although you stopped pulling the piston. This is due to the vacuum created inside and the needle is extremely small to feed the ink quickly. Same way whenever you push the piston, the ink will come out of the syringe slowly and continue to come out of the syringe although you stopped pressing the piston. This is due to the high-pressure inside the syringe and extremely small needle. Infact this is what we can observe in electronic transistors, MOSFETs, vacuum tubes etc dynamically restricting the flow of electrons as well amplifying the electron flow.
network_research_live_demo_kiran_5

So this gives a physical visualization of packet flow in a networking device. Especially cases like networking ports having different bandwidths. As well a case assume you are processing these packets say suppose it is a VPN device, WAN Optimization device, HTTP caching device and so on. The processing of packets adds latency, which means adds resistance to the packet flow. In such cases the Network stack parameters needs to be configured to get optimal performance. Especially the Network packet buffers and so on. And sometimes selecting the right hardware (CPU/RAM/ and so on). So I captured these aspects in a complete video so that my experiments and research may help you to understand these concepts and better architect your Networking components.

Here is the Youtube video link of my complete research:



Suggested Topics:


WAN Optimization and Network Optimization

💎 TOFFEE-MOCHA new bootable ISO: Download
💎 TOFFEE Data-Center Big picture and Overview: Download PDF


おすすめトピック:

Setting up a WAN Emulator within VirtualBox ↗
Saturday' 13-Mar-2021

TOFFEE-Mocha Documentation :: TOFFEE-Mocha-1.0.18-1-x86_64 ↗
Saturday' 13-Mar-2021

First TOFFEE-Mocha Code Release ↗
Saturday' 13-Mar-2021
TOFFEE-Mocha is my dream project which I thought working on it since several years. I want to make a WAN emulation software which is straight forward and simple to use. I used tc scripts along with iptables for testing my TOFFEE (and TrafficSqueezer before TOFFEE) and I am not quite satisfied with the same. As one can understand these scripts are not meant for WAN emulation.

Tracking Live TCP Sessions (connections) - WAN Optimization Device ↗
Saturday' 13-Mar-2021

The TOFFEE Project :: TOFFEE-DataCenter :: WAN Optimization ↗
Saturday' 13-Mar-2021
The TOFFEE Project :: TOFFEE-DataCenter :: Linux Open-Source WAN Optimization

My Lab Battery Purchase and Service logs for Research ↗
Saturday' 13-Mar-2021
Here is a complete log of my lab battery purchase, service record which I maintain in Google drive. These I use for my home (or my family generic use) as well as a part of my home lab. I maintain a detailed log this way to monitor the failure rate of these batteries. This will allow me to select a specific brand/model which has higher success rate and to monitor any premature failure/expiry. The service log helps me to monitor and schedule the next service routine so that I can maintain these batteries in tip-top condition.



First TOFFEE-Mocha Code Release ↗
Saturday' 13-Mar-2021
TOFFEE-Mocha is my dream project which I thought working on it since several years. I want to make a WAN emulation software which is straight forward and simple to use. I used tc scripts along with iptables for testing my TOFFEE (and TrafficSqueezer before TOFFEE) and I am not quite satisfied with the same. As one can understand these scripts are not meant for WAN emulation.

Recording Lab Monthly power-consumption readings for Research ↗
Saturday' 13-Mar-2021
Here is my home lab monthly power-consumption readings for research. This will help to measure and monitor the overall power usage and assess the power requirements. This will help me in future purchases such as UPS, battery upgrades and so on. And as well remove replace old obsolete hardware with new less power-consuming devices.

LoRaWAN - Network Optimization via TOFFEE WAN Optimization ↗
Saturday' 13-Mar-2021
LoRaWAN - Network Optimization via TOFFEE WAN Optimization

Off-Grid Solar Power System for Raspberry Pi ↗
Saturday' 13-Mar-2021
When you choose to use your Raspberry Pi device as your IoT based remote weather station or if you are building Linux kernel (like kernel compilation) within the same, you need a good uninterrupted power source (UPS). But if you are using it on site or in some research camping location you can choose to power your Raspberry Pi device with your custom off-grid solar power source.



Featured Educational Video:
Youtubeで見る - [171//1] 169 Q&A - Add additional HardDrive or storage space in Linux VirtualBox VM ↗

Demo TOFFEE_DataCenter WAN Optimization VM (in VirtualBox) Test Setup ↗
Saturday' 13-Mar-2021
Demo TOFFEE_DataCenter WAN Optimization VM (in VirtualBox) Test Setup

Moon Base and Space Colonization - First we need fast InterPlanetary Internet ↗
Saturday' 13-Mar-2021

iPerf Network Optimization - WAN Optimization Demo ↗
Saturday' 13-Mar-2021

TCP Tune-up and Performance Analysis Graphs - Congestion Control - Research - Dos and Don'ts ↗
Saturday' 13-Mar-2021




Multi-dimensional (Multi-universe) Internet Technology - A Proposal ↗
Saturday' 13-Mar-2021
Currently what we have is a single homogeneous (sort of) WWW Internet. Which we can consider as a single-dimensional network. What I propose is that we can create complete independent multiple Internets with each Internet having its own IP-address space, Domain namespace and an authority to manage Domain names. And these networks/Internets can be entirely IPv4 only based or IPv6 only based.



Research :: Optimization of network data (WAN Optimization) at various levels:
Network File level network data WAN Optimization


Learn Linux Systems Software and Kernel Programming:
Linux, Kernel, Networking and Systems-Software online classes [CDN]


Hardware Compression and Decompression Accelerator Cards:
TOFFEE Architecture with Compression and Decompression Accelerator Card [CDN]


TOFFEE-DataCenter on a Dell Server - Intel Xeon E5645 CPU:
TOFFEE-DataCenter screenshots on a Dual CPU - Intel(R) Xeon(R) CPU E5645 @ 2.40GHz - Dell Server