Meltdown and Spectre impacts on modern compression codecs

[SolidComp]

Hi all – Does anyone know if the Meltdown and Spectre patches slow down compression or decompression in general? I'm not clear on what kinds of operations suffer the most slowdowns. I've heard databases are impacted, and certain kinds of I/O, but I've not dug into the details yet.

Jyrki – Is brotli slower on patched systems? Is Zstd slower? What about something as basic as zlib/gzip?

Joe

[Jyrki Alakuijala]

Hi all – Does anyone know if the Meltdown and Spectre patches slow down compression or decompression in general? I'm not clear on what kinds of operations suffer the most slowdowns. I've heard databases are impacted, and certain kinds of I/O, but I've not dug into the details yet.

Jyrki – Is brotli slower on patched systems? Is Zstd slower? What about something as basic as zlib/gzip?

Joe

I don't know, but I'm very interested to learn about this if anyone can share benchmark results.

[Sportman]

Input 187,074,408,448 bytes - all Bitcoin blockchain blocks till block 516,394 as .tar
Output 152,713,202,154 bytes - botli -1

Virtual Memory = 2 MB
Physical Memory = 3 MB

OS boot single core:

Meltdown and Spectre disabled:

Kernel User Process Global
59.890 613.500 673.390 850.586
62.796 611.000 673.796 848.259
60.843 613.890 674.734 852.859

Meltdown and Spectre enabled:

Kernel User Process Global
73.687 613.593 687.281 878.694
74.312 610.078 684.390 885.106
75.453 611.203 686.656 886.246

------------------------------

OS boot multi core:

Meltdown and Spectre disabled:

Kernel User Process Global
57.187 604.843 662.031 673.131
59.343 609.437 668.781 679.900
57.921 606.796 664.718 675.992

Meltdown and Spectre enabled:

Kernel User Process Global
68.890 606.656 675.546 686.262
68.531 611.375 679.906 699.245
69.109 604.968 674.078 684.374

---------------------------------------------------------------

Input 187,074,408,448 bytes - all Bitcoin blockchain blocks till block 516,394 as .tar
Output 148,029,657,529 bytes - zstd -1

Virtual Memory = 14 MB
Physical Memory = 17 MB

OS boot single core:

Meltdown and Spectre disabled:

Kernel User Process Global
42.140 1198.734 1240.875 1370.639
39.578 1209.593 1249.171 1377.740
39.125 1210.765 1249.890 1379.169

Meltdown and Spectre enabled:

Kernel User Process Global
45.453 1215.265 1260.718 1398.678
47.328 1206.718 1254.046 1394.858
47.687 1196.531 1244.218 1385.758

---------------------------------

OS boot multi core:

Meltdown and Spectre disabled:

Kernel User Process Global
63.687 656.328 720.015 643.378
64.687 655.453 720.140 633.118
66.859 654.281 721.140 632.139

Meltdown and Spectre enabled:

Kernel User Process Global
76.062 656.359 732.421 633.314
74.281 655.250 729.531 648.363
72.546 656.562 729.109 647.317

[Cyan]

Definitely a pretty sensible impact.

Since Meltdown & Spectre mitigation is a kernel topic, I guess most of the extra-load is within kernel-related stuff,
such as threading, task-switching, and most I/O interactions generally.

It would be interesting to know more about the sources of these slow-down,
to understand if user code can do something to reduce triggering them.

[necros]

+1 to know

[SvenBent]

when you disabled the meldtodnw spectra patc h did you do it through the registry or avoide installing the patch all together ?

[Sportman]

when you disabled the meldtodnw spectra patc h did you do it through the registry or avoide installing the patch all together ?

I did not touched the BIOS and OS patches, I used the InSpectre utility https://www.grc.com/inspectre.htm to disable or enable Meltdown and Spectre before reboot.

[Sportman]

Spectre/Meltdown fixes in HPC: Want the bad news or the bad news? It's slower, say boffins:
https://www.theregister.co.uk/2018/0...e_batters_hpc/

[Sportman]

Intel Gags Customers from Publishing Performance Impact of Microcode Updates:
https://www.techpowerup.com/247028/i...rocode-updates

[Sportman]

We have simplified the Intel license to make it easier to distribute CPU microcode updates and posted the new version here:
https://twitter.com/imadsousou/statu...80311753072640
https://01.org/mcu-path-license-2018

Not updated here yet:
https://downloadcenter.intel.com/dow...code-Data-File

[Lucas]

Looks like we can expect some more performance loss now that hyperthreading is considered a security risk.
https://www.theregister.co.uk/2018/0...yperthreading/
https://www.techspot.com/news/75240-...intel-cpu.html

[Sportman]

CPU Utilization is Wrong:
https://opensource.com/article/18/4/...lization-wrong

Video:
https://www.youtube.com/watch?v=QkcBASKLyeU

[Jyrki Alakuijala]

CPU Utilization is Wrong:
https://opensource.com/article/18/4/...lization-wrong

Video:
https://www.youtube.com/watch?v=QkcBASKLyeU

Memory, network and disk/flash can stop the cpu(s) making progress, and particularly so on multi-core systems. Compression can help with all of these, but a compression system is often a memory system hog, and no benchmarking that I know of looks into the memory system load. I believe the impact of the memory system can be 50-100 % speed differences in multiprocessing scaling. For example, two codecs run in A seconds and B seconds, but if you run 10 copies of these codecs in parallel, the other codec runs in A*2 seconds and the other B*4 seconds.

[Sportman]

MDS (Micro-architectural Data Sampling) Tool from RIDL (Rogue In-Flight Data Load) Team:
https://mdsattacks.com/files/mdstool-win.zip
https://mdsattacks.com/files/mdstool-linux.zip
https://github.com/vusec/ridl

[Sportman]

TAA (TSX Asynchronous Abort) and other RIDL (Rogue In-flight Data Load) issues:
https://mdsattacks.com/#ridl-ng
https://mdsattacks.com/files/ridl-addendum.pdf
https://www.youtube.com/watch?v=zaTxBZXE9pQ
https://blogs.intel.com/technology/2...orm-update-ipu
https://software.intel.com/security-...chronous-abort

[Sportman]

Software-based Fault Injection Attacks against Intel SGX (Software Guard Extensions):
https://www.plundervolt.com/doc/plundervolt.pdf
https://www.plundervolt.com/
https://github.com/KitMurdock/plundervolt
https://www.intel.com/content/www/us...-sa-00289.html

[SolidComp]

I was about to post about TSX to see if anyone had found a use for these instructions in compression. It's surprising that they already have security issues – no one seems to be using TSX since they were disabled in many Haswell or Broadwell CPUs.

[Shelwien]

TSX provides a way to restart a chunk of code when a conflict happens during shared memory access: https://gcc.gnu.org/onlinedocs/gcc/x...ntrinsics.html
I suppose it can be useful for some slow MT compressor, like NNCP.
Normally compression doesn't require frequent inter-thread communication because smaller processing units mean worse compression.

[Sportman]

Named Load Value Injection (LVI):
https://www.zdnet.com/article/intel-...w-lvi-attacks/
https://software.intel.com/security-...alue-injection

[Sportman]

The Ongoing CPU Security Mitigation Impact On The Core i9 10900K Comet Lake:
"At least for the workloads tested this round, when booting the new Intel Core i9 10900K "Comet Lake" processor with the software-controlled CPU security mitigations disabled, the overall performance was elevated by about 6% depending upon the workload."
https://www.phoronix.com/scan.php?pa...0k-mitigations

[Sportman]

Straight-Line Speculation (SLS):
https://www.zdnet.com/article/arm-cp...hannel-attack/
https://developer.arm.com/support/ar...-vulnerability
https://developer.arm.com/support/ar...the-whitepaper

[JamesWasil]

Things like this make me wish that Cyrix and a handful of other CPU makers had been able to continue unabated. Even if they needed special cooling to run at increased speeds for performance, it would have been a safe and practical alternative to Intel and AMD CPUs today affected by Spectre and Meltdown.

With more choices besides ARM and alternative (pin compatible) mainstream processors, this might not have been such an issue...

Well, if anyone needs me, I'll be at home on my 80286 6Mhz XT running SCO Xenix and a Lynx browser safely to read Encode.su lol

[Bulat Ziganshin]

the problem isn't on cpu manufacturer side, but in CS itself. any company with modern CPU architecture will have the same problem

[SvenBent]

So I did some testing myself cause i was borred

Windows10 pro Build 18363
Intel I5 4690S

I changed to registry values for memory manages in windows, and after reboot verified meltdown and spectre protectio nwas off with GRC inspectre

GeekBench 5

Patches Off
902 2894
844 2742
852 2763
843 2789
841 2784

Patches ON
914 2891
841 2748
839 2762
843 2767
838 2757


Winsat CPU

Patches off
> CPU LZW Compression 319.47 MB/s
> CPU AES256 Encryption 1851.00 MB/s
> CPU Vista Compression 855.06 MB/s
> CPU SHA1 Hash 2022.84 MB/s
> Uniproc CPU LZW Compression 87.14 MB/s
> Uniproc CPU AES256 Encryption 515.95 MB/s
> Uniproc CPU Vista Compression 234.13 MB/s
> Uniproc CPU SHA1 Hash 564.32 MB/s

> CPU LZW Compression 319.45 MB/s
> CPU AES256 Encryption 1850.78 MB/s
> CPU Vista Compression 855.84 MB/s
> CPU SHA1 Hash 2022.08 MB/s
> Uniproc CPU LZW Compression 88.02 MB/s
> Uniproc CPU AES256 Encryption 515.98 MB/s
> Uniproc CPU Vista Compression 234.96 MB/s
> Uniproc CPU SHA1 Hash 563.92 MB/s

> CPU LZW Compression 319.17 MB/s
> CPU AES256 Encryption 1850.86 MB/s
> CPU Vista Compression 852.82 MB/s
> CPU SHA1 Hash 2023.01 MB/s
> Uniproc CPU LZW Compression 87.85 MB/s
> Uniproc CPU AES256 Encryption 515.97 MB/s
> Uniproc CPU Vista Compression 236.56 MB/s
> Uniproc CPU SHA1 Hash 564.43 MB/s

> CPU LZW Compression 318.79 MB/s
> CPU AES256 Encryption 1851.34 MB/s
> CPU Vista Compression 851.52 MB/s
> CPU SHA1 Hash 2022.77 MB/s
> Uniproc CPU LZW Compression 87.57 MB/s
> Uniproc CPU AES256 Encryption 515.72 MB/s
> Uniproc CPU Vista Compression 237.50 MB/s
> Uniproc CPU SHA1 Hash 564.34 MB/s

> CPU LZW Compression 319.87 MB/s
> CPU AES256 Encryption 1851.17 MB/s
> CPU Vista Compression 855.08 MB/s
> CPU SHA1 Hash 2022.65 MB/s
> Uniproc CPU LZW Compression 87.74 MB/s
> Uniproc CPU AES256 Encryption 516.03 MB/s
> Uniproc CPU Vista Compression 236.99 MB/s
> Uniproc CPU SHA1 Hash 564.21 MB/s


Patches ON
> CPU LZW Compression 319.56 MB/s
> CPU AES256 Encryption 1850.28 MB/s
> CPU Vista Compression 858.72 MB/s
> CPU SHA1 Hash 2021.81 MB/s
> Uniproc CPU LZW Compression 85.56 MB/s
> Uniproc CPU AES256 Encryption 502.77 MB/s
> Uniproc CPU Vista Compression 230.05 MB/s
> Uniproc CPU SHA1 Hash 563.35 MB/s

> CPU LZW Compression 319.35 MB/s
> CPU AES256 Encryption 1849.97 MB/s
> CPU Vista Compression 859.28 MB/s
> CPU SHA1 Hash 2021.58 MB/s
> Uniproc CPU LZW Compression 85.54 MB/s
> Uniproc CPU AES256 Encryption 502.79 MB/s
> Uniproc CPU Vista Compression 229.36 MB/s
> Uniproc CPU SHA1 Hash 549.81 MB/s

> CPU LZW Compression 319.20 MB/s
> CPU AES256 Encryption 1852.14 MB/s
> CPU Vista Compression 858.35 MB/s
> CPU SHA1 Hash 2023.02 MB/s
> Uniproc CPU LZW Compression 86.52 MB/s
> Uniproc CPU AES256 Encryption 515.45 MB/s
> Uniproc CPU Vista Compression 229.39 MB/s
> Uniproc CPU SHA1 Hash 549.85 MB/s

> CPU LZW Compression 318.93 MB/s
> CPU AES256 Encryption 1851.65 MB/s
> CPU Vista Compression 858.41 MB/s
> CPU SHA1 Hash 2022.45 MB/s
> Uniproc CPU LZW Compression 85.50 MB/s
> Uniproc CPU AES256 Encryption 503.07 MB/s
> Uniproc CPU Vista Compression 229.60 MB/s
> Uniproc CPU SHA1 Hash 549.57 MB/s

> CPU LZW Compression 319.79 MB/s
> CPU AES256 Encryption 1852.22 MB/s
> CPU Vista Compression 858.68 MB/s
> CPU SHA1 Hash 2021.78 MB/s
> Uniproc CPU LZW Compression 85.49 MB/s
> Uniproc CPU AES256 Encryption 502.80 MB/s
> Uniproc CPU Vista Compression 233.93 MB/s
> Uniproc CPU SHA1 Hash 563.41 MB/s



Distributed.NET

Patches OFF
OGR-NG: Benchmark for core #2 (cj-asm-sse2)
87,261,778 nodes/sec
87,990,949 nodes/sec
87,539,461 nodes/sec
87,857,098 nodes/sec
87,956,939 nodes/sec

RC5-72: Benchmark for core #4 (YK AVX2)
48,599,183 keys/sec
48,739,659 keys/sec
48,608,994 keys/sec
48,537,309 keys/sec
48,640,024 keys/sec


Patches ON
OGR-NG: Benchmark for core #2 (cj-asm-sse2)
88,450,000 nodes/sec
88,756,774 nodes/sec
88,859,554 nodes/sec
89,011,961 nodes/sec
89,208,522 nodes/sec

RC5-72: Benchmark for core #4 (YK AVX2)
48,969,666 keys/sec
48,737,423 keys/sec
49,042,076 keys/sec
49,241,224 keys/sec
49,122,235 keys/sec

[Sportman]

Input:
66,965,935,529 bytes, BTC 0-xxxx blocks as RAR storage:

Meltdown and Spectre disabled:

Output:
53,410,095,068 bytes, brotli -0:
183.063 sec.
183.010 sec.
183.157 sec.

Output:
51,614,569,992 bytes, zstd -1:
164.479 sec.
165.214 sec.
164.929 sec.

------------------------------------

Meltdown and Spectre enabled:

Output:
53,410,095,068 bytes, brotli -0:
184.210 sec.
184.534 sec.
184.138 sec.

Output:
51,614,569,992 bytes, zstd -1:
166.276 sec.
166.591 sec.
166.774 sec.

[Sportman]

PLATYPLUS - software-based power side-channel attacks (Intel server, desktop and laptop CPUs):
https://platypusattack.com/
https://platypusattack.com/platypus.pdf
https://www.intel.com/content/www/us...-sa-00389.html

[danlock]

It's important to note that although these CPU patches and side-channel attack mitigations are essential for servers, CDNs, and other important WAN-facing machines to apply in order to prevent serious attacks, a typical home user might not need to worry about them. Someone like me, who rarely leaves his PC powered and online for multiple days at a time and never exposes it to the WAN unprotected, can probably get away with leaving those vulnerabilities unpatched (for the most part) and enjoy using the CPU without any slowdowns.

Edit: I forgot about the many laptop users, who don't frequently power down completely, and other home users. They are almost certainly not as careful as I am and probably don't notice a few seconds or a minute added to a task or OS boot time. Those people, the majority, should have as many patches and mitigations enabled as possible... especially when they're employees in a large company. Large companies and organizations are of particular interest to distributors of ransomware, etc.

[hotaru]

It's important to note that although these CPU patches and side-channel attack mitigations are probably essential for servers, CDNs, and other important WAN-facing machines to apply in order to prevent serious attacks, a typical home user will probably not need to worry about them.

these mitigations are essential for anyone who runs untrusted code. that includes typical home users who run JavaScript (much of which is from malicious advertisements) in web browsers.

[Kennon Conrad]

I did some BIOS testing at work about a month ago. The Spectre and Meltdown fixes added almost 1 minute (50%) to the OS boot time.

[danlock]

In case anyone finds it relevant, here are two blog posts from Intel about TSX (Transactional Synchronization Extensions), one of which includes a link to a PDF file containing a Draft Specification of Transaction Language Constructs for C++ (v1.1, 35 pages, 725KB). I'm not sure how useful that C++ spec will be.
Here are the two blog pages:
https://software.intel.com/content/w...explained.html
https://software.intel.com/content/w...n-haswell.html

Chapters 1 & 8 of the Intel Architecture Instruction Set Extensions Programming Reference 319433-012A/Feb 2012 (PDF file, 2.49MB: https://software.intel.com/sites/def.../m/9/2/3/41604 ) contain many details. In fact, all of Chapter 8 (the last chapter, prior to appendices and index), is about TSX! The PDF format makes it easy to click between topics.

*shrug*