Post by Noel on Nov 25, 2015 0:51:22 GMT -6
People sometimes say "Win 10 is Faster".
Faster than what? At doing what?
I did some objective performance testing today, with a fully updated Win 7 system vs. a Win 8.1 system vs. a Win 10 build 10586 system - all with identical provisioning (SSD, 8 GB RAM, 8 cores).
I set out to measure networking and disk performance, as well as raw data crunching power, so I retrieved the same rather large (multi GB) software development baseline on both Win 7 and 10 systems from a local subversion server with no other user activity.
Here's an overview of the times I measured in this "practical usage" software development test. Times in Minutes:Seconds.fraction:
Win 7:
Minutes:Seconds.fraction Activity
12:50.0 Retrieve 52,814 trunk files / folders from SVN
10.77, 2.50, 0.94, 0.90 Batch build 4 projects in GLEW solution
2.97, 2.66, 2.27, 2.27 Batch build 4 projects in LittleCMS solution
1:47.54, 1:39.60, 1:40.93 Batch build 3 projects within the main application
50.0 Delete 19.5 gigabytes of files / folders
Win 8.1:
Minutes:Seconds.fraction Activity
13:45.0 Retrieve 52,814 trunk files / folders from SVN
11.60, 2.73, 1.01, 1.01 Batch build 4 projects in GLEW solution
3.64, 2.43, 2.14, 2.12 Batch build 4 projects in LittleCMS solution
1:47.56, 1:36.70, 1:38.21 Batch build 3 projects within the main application
30.0 Delete 19.5 gigabytes of files / folders
Win 10 build 10586:
Minutes:Seconds.fraction Activity
15:31.0 Retrieve 52,814 trunk files / folders from SVN
10.79, 2.71, 1.04, 0.98 Batch build 4 projects in GLEW solution
3.29, 2.54, 2.28, 2.23 Batch build 4 projects in LittleCMS solution
2:30.78, 1:34.65, 1:35.34 Batch build 3 projects within the main application
47.0 Delete 19.5 gigabytes of files / folders
Notably Windows 10 DID slightly outpace Windows 7 in a couple of the above areas, and in DELETING the whole file/folder structure in the end. BUT... Windows 7 did the whole job, soup to nuts, in about 3 minutes less time overall.
Honestly, I believe based on what I'm seeing above, that this represents a performance improvement in Win 10 build 10586 over what I have measured before in 10240.
Reflecting on the above test, and in all fairness and openness, there's a factor out of my control that could contribute to small variances in the individual product build times: There are code signing steps, which require confirmation from an online time server. Based on what I saw going by the differences wouldhave been only fractions of a second - I didn't notice the process stall, BUT... One of the projects in the main application DID take a LOT longer for Windows 10 than 7.
This is actually encouraging for Windows 10's disk I/O performance, because it implies Microsoft may have optimized the file system some in 10586, though it implies that at least some parts of file system access (writes?) or networking are still slower in Windows 10.
More testing
Ask anyone who thinks Win 10 is faster than Win 7 to select all files in the root folder of C:\ and do Properties, timing how long it takes Explorer to enumerate them all. Ask them to do it twice, to test performance when the disk is involved vs. when the data is all cached in RAM.
Same identically provisioned systems as above, freshly booted and allowed to settle for a few minutes, then Explorer instructed to count up all files on the SSD.
Windows 7: 243,214 files in 67,780 folders
Seconds Activity
17.6 First time (13,819 files per second straight off the SSD)
4.4 Second time (55,276 files per second cached)
Windows 8.1: 263,458 files in 44,584 folders
Seconds Activity
41.2 First time ( 6,395 files per second straight off the SSD)
19.6 Second time (13,442 files per second cached)
Windows 10: 214,567 files in 44,784 folders
Seconds Activity
47.4 First time ( 4,527 files per second straight off the SSD)
28.0 Second time ( 7,663 files per second cached)
Since the amount of time needed to count up files in a Properties dialog isn't all THAT important, and I don't want to hinge on just a single Explorer deficiency, I decided to do a different operation that traverses the file system. I searched the entire filesystem for a partial wildcard filename from the root using: DIR C:\*srgb*.* /S /B
Windows 7: 243,214 files in 67,780 folders
Seconds Activity
25 First time ( 9,700 files per second straight off the SSD)
8 Second time (30,000 files per second cached)
Windows 8.1: 263,458 files in 44,584 folders
Seconds Activity
16 First time (16,000 files per second straight off the SSD)
5 Second time (50,000 files per second cached)
Windows 10: 214,567 files in 44,784 folders
Seconds Activity
26 First time ( 8,300 files per second straight off the SSD)
10 Second time (21,000 files per second cached)
The DIR command traversing subdirectories in Win 8.1 outpaces both Win 7 and 10, with 10 being slowest.
File system operations are at the core of everything. The faster they go, the faster the system seems to respond, and they're clearly slower on Windows 10. And I don't know about you, but I do use Explorer a lot. Hence my deep skepticism at claims that Windows 10 "is faster".
It strikes me that one reason so many people haven't flocked to Windows 10 yet is that not everyone blindly believes the hype that it's faster.
Some specific network performance testing
I have a tool called filecrc I wrote a long while back that just enumerates files and folders and "adds up" all the bytes using a CRC-32 algorithm. Since the computing isn't very complex, this is good for measuring raw I/O speed.
If I use it to read nearly a gigabyte of data from a machine across the LAN (running Win 7), I'm thinking it'll give us an idea of how efficient Windows Networking is. Though many folks may not use a LAN, this will be important to folks who do. I saw a surprising reversal of performance figures (and yes, I repeated the test several times with the same results).
Win 7: 23 seconds:
Win 8.1: 14 seconds:
Win 10: 16 seconds:
Surprisingly, Windows 10 outpaced Windows 7 at reading files across my LAN network - to another Win 7 machine, but it still doesn't quite match the speed of Win 8.1.
I encourage you to add your observations / measurements / benchmark results to this thread.
-Noel