Download TCSS 422: Operating Systems Course Material and Objectives and more Lecture notes Operating Systems in PDF only on Docsity!
School of Engineering and Technology
June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma Beyond Physical Memory, I/O Devices, HDDs
Wes J. Lloyd
School of Engineering and Technology
University of Washington - Tacoma TCSS 422: OPERATING SYSTEMS Andrew Oberhardt Senior Software Engineer, Axon Company Alumni Virginia Tech Previously: Software Engineer at Microsoft (10+ years), Code.org, and Amazon Axon Company
https://www.axon.com/company
Open positions, hiring
GUEST SPEAKER, Q&A June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18. Questions from 5/ Tutorial 2 (pthreads, locks, conditions) – due Thurs June 4 Quiz 3 posted – Active Reading Ch. 19 – due Tues June 2 Assignment 3 – on Linux kernel programming – of fered in “tutorial” format – due Sat June 13 Quiz 4 – Page Tables Practice Final Exam Questions – Thursday – 2nd^ hour Chapter 22: Cache Replacement Policies: Workl oad E x a mple s Chapter 36: I/O Devices Polling vs. Interrupts, Programmed I/O, Direct Memory Access Chatper 37: Hard Disk Drives June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
OBJECTIVES – 6/
Please classify your perspective on material covered in today’s class (38 respondents): 1-mostly review, 5-equal new/review, 10-mostly new Average – 7.3 ( from 6.62) Please rate the pace of today’s class: 1-slow, 5-just right, 10-fast Average – 5.83 ( from 5.84) June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
MATERIAL / PACE
June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
FEEDBACK FROM 5/
Questions from 5/ Tutorial 2 (pthreads, locks, conditions) – due Thurs June 4 Quiz 3 posted – Active Reading Ch. 19 – due Tues June 2 Assignment 3 – on Linux kernel programming – of fered in “tutorial” format – due Sat June 13 Quiz 4 – Page Tables Practice Final Exam Questions – Thursday – 2nd^ hour Chapter 22: Cache Replacement Policies: Wo rkl oad E x a mple s Chapter 36: I/O Devices Polling vs. Interrupts, Programmed I/O, Direct Memory Access Chatper 37: Hard Disk Drives June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
OBJECTIVES – 6/
School of Engineering and Technology
Questions from 5/ Tutorial 2 (pthreads, locks, conditions) – due Thurs June 4 Quiz 3 posted – Active Reading Ch. 19 – due Tues June 2 Assignment 3 – on Linux kernel programming – of fered in “tutorial” format – due Sat June 13 Quiz 4 – Page Tables Practice Final Exam Questions – Thursday – 2nd^ hour Chapter 22: Cache Replacement Policies: Workl oad E x a mple s Chapter 36: I/O Devices Polling vs. Interrupts, Programmed I/O, Direct Memory Access Chatper 37: Hard Disk Drives June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
OBJECTIVES – 6/
Questions from 5/ Tutorial 2 (pthreads, locks, conditions) – due Thurs June 4 Quiz 3 posted – Active Reading Ch. 19 – due Tues June 2 Assignment 3 – on Linux kernel programming – of fered in “tutorial” format – due Sat June 13 Quiz 4 – Page Tables Practice Final Exam Questions – Thursday – 2nd^ hour Chapter 22: Cache Replacement Policies: Wo rkl oad E x a mple s Chapter 36: I/O Devices Polling vs. Interrupts, Programmed I/O, Direct Memory Access Chatper 37: Hard Disk Drives June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
OBJECTIVES – 6/
Questions from 5/ Tutorial 2 (pthreads, locks, conditions) – due Thurs June 4 Quiz 3 posted – Active Reading Ch. 19 – due Tues June 2 Assignment 3 – on Linux kernel programming – of fered in “tutorial” format – due Sat June 13 Quiz 4 – Page Tables Practice Final Exam Questions – Thursday – 2nd^ hour Chapter 22: Cache Replacement Policies: Workl oad E x a mple s Chapter 36: I/O Devices Polling vs. Interrupts, Programmed I/O, Direct Memory Access Chatper 37: Hard Disk Drives June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
OBJECTIVES – 6/
Questions from 5/ Tutorial 2 (pthreads, locks, conditions) – due Thurs June 4 Quiz 3 posted – Active Reading Ch. 19 – due Tues June 2 Assignment 3 – on Linux kernel programming – of fered in “tutorial” format – due Sat June 13 Quiz 4 – Page Tables Practice Final Exam Questions – Thursday – 2nd^ hour Chapter 22: Cache Replacement Policies: Wo rkl oad E x a mple s Chapter 36: I/O Devices Polling vs. Interrupts, Programmed I/O, Direct Memory Access Chatper 37: Hard Disk Drives June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
OBJECTIVES – 6/
CHAPTER 21/22: BEYOND PHYSICAL MEMORY June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18. Questions from 5/ Tutorial 2 (pthreads, locks, conditions) – due Thurs June 4 Quiz 3 posted – Active Reading Ch. 19 – due Tues June 2 Assignment 3 – on Linux kernel programming – of fered in “tutorial” format – due Sat June 13 Quiz 4 – Page Tables Practice Final Exam Questions – Thursday – 2nd^ hour Chapter 22: Cache Replacement Policies: Wo rkl oad E x a mple s Chapter 36: I/O Devices Polling vs. Interrupts, Programmed I/O, Direct Memory Access Chatper 37: Hard Disk Drives June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
OBJECTIVES – 6/
School of Engineering and Technology
No-Locality (Random Access) Workload Perform 10,000 random page accesses Across set of 100 memory pages June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
WORKLOAD EXAMPLES: NO-LOCALITY
When the cache is large enough to fit the entire workload, it doesn’t matter which policy you use. 80/20 Workload Perform 10,000 page accesses, against set of 100 pages 80% of accesses are to 20% of pages (hot pages) 20% of accesses are to 80% of pages (cold pages) June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
WORKLOAD EXAMPLES: 80/
LRU is more likely to hold onto hot pages (recalls history) Looping sequential workload Refer to 50 pages in sequence: 0, 1, …, 49 Repeat loop June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
WORKLOAD EXAMPLES: SEQUENTIAL
Random performs better than FIFO and LRU for cache sizes < 50 Algorithms should provide “scan resistance” June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.2 3 Implementing last recently used (LRU) requires tracking access time for all system memory pages Times can be tracked with a list For cache eviction, we must scan an entire list Consider: 4GB memory system (2^32 ), with 4KB pages (2^12 ) This requires 2^20 comparisons !!! Simplification is needed Consider how to approximate the oldest page access June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
IMPLEMENTING LRU
School of Engineering and Technology
Harness the Page Table Entry (PTE) Use Bit
HW sets to 1 when page is used
OS sets to 0
Clock algorithm (approximate LRU)
Refer to pages in a circular list
Clock hand points to current page
Loops around
IF USE_BIT=1 set to USE_BIT = 0 IF USE_BIT=0 replace page June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
IMPLEMENTING LRU - 2
Not as efficient as LRU, but better than other replacement algorithms that do not consider history June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
CLOCK ALGORITHM
Consider dirty pages in cache
If DIRT Y (modified) bit is FALSE
No cost to evict page from cache
If DIRT Y (modified) bit is TRUE
Cache eviction requires updating memory
Contents have changed
Clock algorithm should favor no cost eviction
June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
CLOCK ALGORITHM - 2
On demand demand paging Prefetching Preload pages based on anticipated demand Prediction based on locality Access page P, suggest page P+1 may be used
What other techniques might help anticipate required
memory pages?
Prediction models, historical analysis In general: accuracy vs. effort tradeoff High analysis techniques struggle to respond in real time June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
WHEN TO LOAD PAGES
Page swaps / writes
Group/cluster pages together
Collect pending writes, perform as batch
Grouping disk writes helps amortize latency costs
Thrashing
Occurs when system runs many memory intensive
processes and is low in memory
Everything is constantly swapped to-and-from disk
June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
OTHER SWAPPING POLICIES
Working sets
Groups of related processes
When thrashing: prevent one or more working
set(s) from running
Temporarily reduces memory burden
Allows some processes to run, reduces thrashing
June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
OTHER SWAPPING POLICIES - 2
School of Engineering and Technology
Consider an arbitrary canonical “standard/generic” device Two primary components Interface (registers for communication) Internals: Local CPU, memory, specific chips, firmware (embedded software) June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
CANONICAL DEVICE
Status register Maintains current device status Command register Where commands for interaction are sent Data register Used to send and receive data to the device June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
CANONICAL DEVICE:
HARDWARE INTERFACE
General concept: The OS interacts and controls device behavior by reading and writing the device registers. Common example of device interaction June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
OS DEVICE INTERACTION
Poll- Is device available? Poll – Is device done? Command parameterization Send command OS checks if device is READY by repeatedly checking the STATUS register Simple approach CPU cycles are wasted without doing meaningful work Ok if only a few cycles, for rapid devices that are often READY BUT polling, as with “spin locks” we understand is inefficient June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
POLLING
For longer waits, put process waiting on I/O to sleep Context switch (C/S) to another process When I/O completes, fire an interrupt to initiate C/S back Advantage: better multi-tasking and CPU utilization Avoids: unproductive CPU cycles (polling) June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
INTERRUPTS VS POLLING
Interrupts are not always the best solution How long does the device I/O require? What is the cost of context switching? June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
INTERRUPTS VS POLLING - 2
If device I/O is fast polling is better. When I/O time < 1 CPU time slice (e.g. 10 ms) If device I/O is slow interrupts are better. When I/O time > 1 CPU time slice What is the tradeoff space?
School of Engineering and Technology
Alternative: two-phase hybrid approach Initially poll, then sleep and use interrupts Issue: livelock problem Common with network I/O Many arriving packets generate many many interrupts Overloads the CPU! No time to execute code, just interrupt handlers! Livelock optimization Coalesce multiple arriving packets (for different processes) into fewer interrupts Must consider number of interrupts a device could generate June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
INTERRUPTS VS POLLING - 3
TCSS 422 WILL RETURN AT ~2:45PM June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L15.
To interact with a device we must send/receive
DATA
There are two general approaches:
Programmed I/O (PIO):
Port mapped I/O (PMIO)
Memory mapped I/O (MMIO)
Direct memory access (DMA)
June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
DEVICE I/O
TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma From https://en.wikipedia.org/wiki/Parallel_ATA I/O performed on the CPU CPU time is consumed performing I/O CPU supports data movement (input/output) PIO is slow: CPU is occupied with meaningless work June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
PROGRAMMED I/O (PIO)
PIO
Legacy serial ports Legacy parallel ports PS/2 keyboard and mouse Legacy MIDI, joysticks Old network interfaces June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
PIO DEVICES
School of Engineering and Technology
Layers of I/O abstraction in Linux C functions (open, read, write) issue block read and w rite requests to the generic block layer June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
FILE SYSTEM ABSTRACTION
Too much abstraction Many devices provide special capabilities Example: SCSI Error handling SCSI devices provide extra details which are lost to the OS Buggy device drivers 70% of OS code is in device drivers Device drivers are required for every device plugged in Drivers are often 3rd^ party, which is not quality controlled at the same level as the OS (Linux, Windows, MacOS, etc.) June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
FILE SYSTEM ABSTRACTION ISSUES
CH. 37: HARD DISK DRIVES June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18. Questions from 5/ Tutorial 2 (pthreads, locks, conditions) – due Thurs June 4 Quiz 3 posted – Active Reading Ch. 19 – due Tues June 2 Assignment 3 – on Linux kernel programming – of fered in “tutorial” format – due Sat June 13 Quiz 4 – Page Tables Practice Final Exam Questions – Thursday – 2nd^ hour Chapter 22: Cache Replacement Policies: Wo rkl oad E x a mple s Chapter 36: I/O Devices Polling vs. Interrupts, Programmed I/O, Direct Memory Access Chatper 37: Hard Disk Drives June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
OBJECTIVES – 6/
Chapter 37 HDD Internals Seek time Rotational latency Transfer speed Capacity Scheduling algorithms June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
OBJECTIVES
Primary means of data storage (persistence) for decades Remains inexpensive for high capacity storage 2020: 16 TB HDD - $400, ~15.3 TB SSD - $4, Consists of a large number of data sectors Sector size is 512-bytes An n sector HDD can be is addressed as an array of 0..n-1 sectors June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
HARD DISK DRIVE (HDD)
School of Engineering and Technology
Writing disk sectors is atomic (512 bytes) Sector writes are completely successful, or fail Many file systems will read/write 4KB at a time Linux ext3/4 default filesystem blocksize – 4096
Same as typical memory page size
June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
HDD INTERFACE
mkefs.ext4 -i <bytes-per-inode>
Formats disk w/ ext4 filesys with specified byte-to-inode ratio Today’s disks are so large, some use cases with many small files can run out of inodes before running out of disk space Each inode record tracks a file on the disk Larger bytes-per-inode ratio results in fewer inodes Default is around ~ Value shouldn't be smaller than blocksize of filesystem Note: It is not possible to expand the number of inodes after the filesystem is created, - be careful deciding the value Check inode stats: tune2fs –l /dev/sda1 ( d isk dev na m e) June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
BLOCK SIZE IN LINUX EXT
Host ~2,000,000 small XML files totaling 9.5 GB on a ~20GB filesystem on a cloud-based Virtual Machine With default inode ratio (4096 block size), only ~488,000 files will fit Drive less than half full, but files will not fit! HDDs support a minimum block size of 512 bytes OS filesystems such as ext3/ext4 can support “finer grained” management at the expense of a larger catalog size Small inode ratio- inodes will considerable % of disk space June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
EXAMPLE: USDA SOIL EROSION MODEL
WEB SERVICE (RUSLE2)
Free space in bytes (df)
Device total size bytes- used bytes- free usage /dev/vda2 13315844 9556412 3049188 76% /mnt
Free inodes (df –i) @ 512 bytes / node
Device total inodes used free usage /dev/vda2 3552528 1999823 1552705 57% /mnt June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
EXAMPLE: USDA SOIL EROSION MODEL
WEB SERVICE (RUSLE2) - 2
Torn write When OS uses larger block size than HDD Block writes not atomic - they SPAN multiple HDD sectors Upon power failure only a portion of the OS block is written – can lead to data corruption… HDD access Sequential reads of sectors is fastest Random sector reads are slow Disk head continuously must jump to different tracks June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
HDD INTERFACE - 2
June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
HDD PLATTER
Made from aluminum coated with thin magnetic layer HDD records on both sides of each platter Data is stored by inducing magnetic changes
School of Engineering and Technology
Acceleration coasting deceleration settling Acceleration: the arm gets moving Coasting: arm moving at full speed Deceleration: arm slow down Settling: Head is carefully positioned over track Settling time is often high, from .5 to 2ms June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
FOUR PHASES OF SEEK
Data transfer
Final phase of I/O: time to read or write to disk
surface
Complete I/O cycle:
1. Seek (accelerate, coast, decelerate, settle)
2. Wait on rotational latency (until track aligns)
3. Data transfer
June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
HDD I/O
Sectors are offset across tracks to allow time for head to reposition for sequential reads Without track skew, when head is repositioned sector would have already been passed June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
TRACK SKEW
June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
TRACK SKEW - 2
Buffer to support caching reads and writes Improves drive response time Up to 256 MB, slowly have been growing Two styles Writeback cache Report write complete immediately when data is transferred to HDD cache Dangerous if power is lost Writethrough cache Reports write complete only when write is physically completed on disk June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
HDD CACHE
Can calculate I/O transfer speed with: I/O Time: Ttransfer = DATAsize x RateI/O Rate of I/O: June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
TRANSFER SPEED
School of Engineering and Technology
Compare two disks:
- Random workload: 4KB (random read on HDD)
- Sequential workload: 100MB (contiguous sectors)
Calculate Trotation from rpm (rpmrps, time for 1 rotation / 2) June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
EXAMPLE: I/O SPEED
- Random workload: 4KB (random read on HDD)
- Sequential workload: 100MB (contiguous sectors) June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
EXAMPLE: I/O SPEED
There is a huge gap in drive throughput between random and sequential workloads Ttransfer = Datasize x RateI/O 4 KB 100 MB See sample HDD configurations here: Up to 20 TB https://www.westerndigital.com/products/data-center- drives#hard-disk-hdd June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
MODERN HDD SPECS
Disk scheduler: determine how to order I/O requests Multiple levels - OS and HW OS: provides ordering HW: further optimizes using intricate details of physical HDD implementation and state June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
DISK SCHEDULING
Disk scheduling – which I/O request to schedule next Shortest Seek Time First (SSTF) Order queue of I/O requests by nearest track June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
SSTF – SHORTEST SEEK TIME FIRST
Problem 1: HDD abstraction Drive geometry not available to OS. Nearest-block-first is a comparable alternate algorithm. Problem 2: Starvation Steady stream of requests for local tracks may prevent arm from traversing to other side of platter June 2, 2020 TCSS422: Operating Systems [Spring 2020] School of Engineering and Technology, University of Washington - Tacoma L18.
SSTF ISSUES
- Slide
- WJL1 Wes J. Lloyd, 5/30/
