%include "default.mgp"
%default 1 bgrad 4 4 4 76 1 "black" "blue"
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%page
%nodefault
%center, size 7, font "standard", fore "white", vgap 20

Disk And Storage Changes

%center, size 4, font "standard", fore "white", vgap 20

Alan Cox
%page

In The Beginning

MFM/RLL controllers
	Interleave
	Register command set
	Sector at a time/polled
	250-500K/sec
	C/H/S addressing - requires OS mapping
	30-100MB, 4MB RAM, 1K OS block sizes

%page

In The Beginning

Early ISA SCSI
	Dumb adapters
	Smart adapters
	Scatter Gather limits
	ISA 16Mb limit
	Disconnect

%page

Early IDE

IDE
	Processor driven
	But we did finally get IRQs
	Forward MFM/RLL controller access cycles down cable to disk
	Saved electronics/money
	Moved the intelligence

%page

ATA Standards

ATA
	Rapid expansion in disk size
	Logical Block Addressing
	Rename registers to LBA bits
	Initial DMA support
	VLB makes faster modes useful
	No programming standards

%page

ATA on PCI

PCI SFF8018
	Unifies basic IDE on PCI functionality
	Unifies basic DMA with BIOS mode setting

From Disk To Device
	Functionality evolving from read/write I/O device
	Door locking
	Feature reporting
	Identify

%page

Limits And Usage Changes

Limits
	Tied to physical C/H/S assumptions still

Bigger I/O
	Multi sector I/O with one interrupt
	4K+ I/O becomes a bigger win
	Block sizes of file systems now often 4K
	Evolving Linux page cache favours 4K I/Os

%page

ATA4

ATAPI
	SCSI type commands for CD to ATA  
	Causes older chips headaches
	SCSI layer ide-scsi and a translator
	Alternative ide-cd driver

Caches
	Caches becoming common
	OS needs to flush them

%page

Ultra DMA

Ultra DMA
	CRC error handling
	Mode changes can now be required after boot. 
		Makes a mess
	Data transfer is fast
	Disk read rate is fast
	Command issue time is slow
	Seeks still take forever

%page

Taskfiles - ATA as messages

Taskfiles
	Redefine the registers as a message
		In serial ATA it *is* a message
	ACPI suspend/restore use taskfiles
	drivers/ide talks taskfiles ..ish
	Command set evolving
		SMART, HPA, Locking
	Still no solution to seek or command rate

%page

LBA48 - Large Disk

LBA48
	Two sets of writes to each register
	Bigger disk
	Bigger I/O size
	Bigger controller breakage
	Even slower command issue rates
	Bigger I/O size rarely useful

%page

Serial ATA

Serial ATA
	Moves away from SFF-8018
	64bit addressing for DMA
	Multiple outstanding commands - NCQ
	Controllers all have differing I/O models
	Hotplug at disk and controller level
	AHCI becoming new standard model

Attempts to retrofit bits to drivers/ide - not pretty
New libata layer uses the block/scsi layer work

%page

Higher Level Happenings

Problems
	Databases want ordering/storage guarantees
	Journalling file systems
	SCSI layer starts using TCQ
	Not much older ATA can do or needed to

Multiple Outstanding Commands
	Which commands
	How do we merge them
	How do we be fair
	What about ordering for journals

%page

Barriers and Ordering

Barriers
	2.6 evolving explicit barrier functionality
	ATA can use this with cache flush commands
	SCSI can use TCQ
	Error handling for ordered commands is hard

%page

Pluggable schedulers

Pluggable schedulers
	Different tasks have different needs
		Latency v fairness v throughput
	Disk improvements low relative to CPU gains
	Seek time is little improved in years (5-6x)
	Disk throughput has risen 80x

	We now have more CPU and more need to be smart

%page

Anticipatory Scheduler

Anticipatory Scheduler
	Classic one way elevator
	Adds opportunistic _short_ back seeks
	FIFO expiry times to avoid starvation
	Read and write batching when possible
	Tries to learn "think v I/O" time of apps
		Batch I/O better by delaying
	Good for general I/O
	Dire for swap

%page

Deadline Scheduler

Deadline Scheduler
	Similar to anticipatory
	Focus heavily on deadline hitting
	Don't delay I/O for batching
	Optionally do minimal merging for smart subsystems
	Good for databases

%page

CFQ and No-Op

CFQ (Completely Fair Queuing)
	Draws on the CFQ ideas in networking
	Try and provide balanced bandwidth to processes
	Oriented to desktops
	CPU and memory usage to manage is higher
	Fairness costs throughput

No-Op
	For smart storage systems
	Minimise CPU and memory cost
	Minimise latency
	Simply vomit it all directly at the disk controller

%page

Unsolved Problems

Application smartness
	Opening millions of files is expensive
	Random gratuitous seeking is costly too
	Gtk/gnome startup

Research problems
	Re-ordering binaries so they page in linearly
	Log structured swap
	Anticipatory swap-in

%page

Storage Has Changed Dramatically

Historically
	Simple algorithms to save CPU
	Tightly linked to disk
	OS could understand disk layout

Today
	Disk is a storage appliance
	CPU power is cheap
	I/O is a bottleneck
	Disk seek time has not scaled

Need for more smarter algorithms.
May all change again with new technology

%page

%page