Reflection on OSI model
In the world of inter-networking, the most popular standard has to be TCP/IP protocol, which forms the basis of the Internet connecting billions of computers and devices around the world. It’s so popular and vital to our daily life now a days that it’s used even when communication is local across compatible networks. For example, popular network file system NFS uses IP protocol even though it’s used mainly in homogenous ethernet LAN environment.
The software standards like TCP/IP allowing reliable communication without demanding reliable networks is the key enabling technologies for inter-networking. Looking back the standard battle with ATM in 1990s, one key success factor is it’s decomposed layered hierarchy. Each layer taking responsibility for just a portion of the overall communication task. To avoid ambiguous in terms, Open System Interconnect (OSI) model is developed which describes networks as a series of layers:
| Layer number | layer name | Main function | Example protocol |
|---|---|---|---|
| 7 | Application | Application running over network | DNS, Web |
| 6 | Presentation | Translate between Application and network format | |
| 5 | Session | Session lifecycle management across network | RPC,named pipes |
| 4 | Transport | Additional connection below session layer | TCP |
| 3 | Network | Translate between logic address/name to physical address | IP |
| 2 | Data Link | Transform between packet and bits | Ethernet |
| 1 | Physical | TX/RX bit stream over physical media | IEEE 802 |
The key to protocol families following OSI model is that communication occurs logically at same level of the protocol between sender and receiver, while services from lower level implement it. Just like abstract data types which simplifies the programmer’s task by hiding the implementation details of the data type, offering services needed by upper layer at each protocol layer makes the standard easier to understand and implement with cross industry compatibility, especially for lower level protocol vendors.
For 20 years since mid 1990s, Internet has enjoyed exponential growth thanks to the dominant position of TCP/IP. Network vendors like Cisco took lead and protocols are added and expanded with a bottom up approach to protect the investment in lower level infrastructures. It is until recent years we see some challenges and movements against long-standing TCP/IP stack and traditional network infrastructures (both topology and component levels), driven by OTT providers’ performance, scalability and fault tolerance demands.
The analysis for the huge success of AWS, Azure, GCP and Ali cloud in the past 5-10 years deserves a dedicate study series but here is one important factor from technology point of view: they just need purposely built network for their intent business need, so they built proprietary homogeneous network infrastructure from scratch with optimized and simplified protocol stack, achieving cost, power consumption and networking efficiency.
It is on the right track but start to show limitations facing challenges from:
- Data volume and dispersal driven by AI and IoT
- Dynamic business logic driven by tenant IT migration
Simply put, the “simplification” approach OTT providers have enjoyed for their past success is just not enough when facing these new challenges, which demands REAL application driven design, with Top Down principle.
Back to OSI model, we have to abandon the assumption of TCP/IP, Ethernet, etc and go back to basic:
- Layers are for separation of responsibilities in a hierarchical way
- Upper layer should not aware or concern about lower layer
- Lower layer is responsible of fulfilling or rejecting the requests from upper layer
Which means, instead of the bottom-up approach where upper layer operates based on what lower level offers, upper layer operates purely based on its own need and send only intent to lower layer, where lower layer has the flexibility to choose available and viable technology to fulfill or reject the request.
Think of a future where we can have harmonized heterogeneous network clusters (devices, topologies or protocols) within data centers and extend into edge, campus even SoC, tailored for different compute-storage usage, totally transparent to applications layer; Think of a laptop where all bandwidth of wired and available WiFi band can be utilized simultaneously; Think of google-search-like information query from application; Think of instructions rather than data been networked to fulfill the request using remote data, etc.
This is a challenging time for networking, but it’s also a perfect time for breaking away: Welcome to an era of converged compute, storage and networking, friend!