Unbuffered I/O Problem

The problem

• Kernel issues read command
• But then…process gets swapped out…
• …leaving the I/O operations blocked while waiting for memory
• This situation can lead to a deadlock scenario, where the process is unable to proceed due to the blocked I/O operation

I/O Buffering

• Use kernel memory to buffer data
  • i.e., temporarily store data being transferred between devices and processes.
• Allows efficient data handling by enablind read-ahead operations
  • Fetches additional data in anticipation of future requests
  • Reduce delays in data retrieval

Single Buffering

• Transfer data to kernel buffer
• Then move it to userspace
• Requires kernel buffer to be available until data is copied to userspace
  • Potentially causes delays id buffer is occupied

Double Buffer

• Transfer data to one kernel buffer while…
• …simultaneously copying another buffer to userspace
  • Allows increases request servicing capabilities
  • can enhance performance by enabling read-ahead operations

Ring Buffer

• Stores data in a fixed-size buffer with a wrap-around mechanism
• Good for data streams where the oldest data is overwritten by newest data when buffer is full
• Efficient for handling bursty I/O patterns

Buffering Speedup

• Performance improvement achieved by using buffers to optimise data transfer and processing
  • $T$: Time required to fetch input block
  • $C$: Computation time between input requests
  • $M$: Time to move a block from buffer to userspace
• Speedup by reducing the time required to fetch input blocks and moving data between buffers and userspace
• Execution Time:
  • No Bufffer: $T+C$
  • Single Buffer: $\text{max}(C, T) + M$
  • Double Buffer: $\text{max}(C, T)$

Zero Copy I/O

• What’s better than lots of buffering? No buffering!!!
• Eliminates the need for data buffering during I/O
  • Reduces latency by directly transferring data between applications and devices without intermediate buffering stages
  • Requires driver support and memory that isn’t swapped out