ZSTD Compression

[VincentJ]

I have more results in progress on the RAMdisk. Hopefully i’ll get them done tonight.

[Quazz]

@VincentJ From what I’ve read this compression algorithm is targetted specifically at modern CPUs, leveragering their instructions sets, which older CPUs will lack.

Meaning that tests will be needed on older hardware, as there will be people using it for quite some time. If there’s a huge time penalty there, it will still be a no go for a lot of people.

[Junkhacker]

ok, here are my pseudo scientific results:

pigz vs pzstd (parallel implementation of zstandard, experimental)
tests were performed using a windows 7 image
(larger of 2 partitions only)
uncompressed image file size:34650439624

(de)compression tests were performed as closely as i could to emulate fog’s operation methods (without doing too much work :P). files were cat-ed from a mounted nfs share and piped into the programs, with results saved to an SSD. test machine was running Lubuntu instead of a custom FOS init, because i’m lazy.

pigz -6 compression
duration: 6:06
file size: 17548659028

pigz -6 decompression
duration: 6:00

pzstd default (3?) compression
duration: 5:16
file size: 16967988207

pzstd decompression default compression file
duration: 3:17

pzstd -6 compression
duration: 6:11
file size: 16247155611

pzstd decompression -6 compression file
duration: 3:16

pzstd -9 compression
duration: 10:00
file size: 16084180231

pzstd decompression -9 compression file
duration: 3:21

Edited to add zst compression level 6

[Quazz]

@Junkhacker Interesting, what kind of CPU did the test device have?

[Junkhacker]

@Quazz it’s an Optiplex 3020. i5

it may be worth pointing out that the pigz performance from my tests is not really representative of what i typically see when actually fog imaging this machine. testing was faster at compression by a significant amount, and a bit slower at decompression, than my experience. not sure what that says about the usefulness of these tests.

[VincentJ]

SO…

16,390,624KB file removed from the compressed windows 10 image.
on a 34GB RAMdisk.

Copying the file on the RAMdisk is 740MB/s so that is well above what we need for imaging for most people.

Lets try some things to get some numbers.

Compression - Compressed size - Compression time - Decompression time
zstd lvl1 - 7,940,779KB - 50 seconds - 38 seconds
zstd lvl3 - 7,420,268KB - 75 seconds - 40 seconds
zstd lvl5 - 7,286,951KB - 128 seconds - 40 seconds
zstd lvl8 - 7,070,670KB - 261 seconds - 41 seconds
zstd lvl11 - 6,967,155KB - 425 seconds - 41 seconds
zstd lvl14 - 6,942,360KB - 674 seconds - 42 seconds
zstd lvl17 - 6,781,375KB - 1,618 seconds - 42 seconds
zstd lvl20 - 6,471,945KB - 2,416 seconds - 43 seconds
zstd lvl22 - 6,214,702KB - 3,970 seconds - 45 seconds

pigz.exe --keep -0 a:\d1p2 - 16,393,125KB - 72 seconds - 80 seconds
pigz.exe --keep -3 a:\d1p2 - 7,783,303KB - 292 seconds - 158 seconds (157 seconds)
pigz.exe --keep -6 a:\d1p2 - 7,535,149KB - 518 seconds - 149 seconds
pigz.exe --keep -9 a:\d1p2 - 7,512,046KB - 1,370 seconds - 149 seconds

Windows 10 Pro, 4 vCPU 42GB RAM with 34GB RAM Disk.
Host XenServer 7.0, Dual E5-2603 v3, 64GB RAM, HDD Raid 1.
Other VMs moved to the other hosts in the pool.

Decompression seems to not use all CPU with PIGZ… around 50%…
Compression does use all 100% CPU
Decompression with zstd does use all CPU - but most were around 400MB/s so possibly I’m hitting some other limit.

[VincentJ]

So, both of us have results that show zstd decompressing quicker and having better ratios.

I’m going to reconfigure my VM to only have 1 vCPU at 1.6GHz to see if i can get more useful decompression results.

I redid the pigz -3 decompression test twice to confirm it was slower than the others… Not what i was expecting but that is what happened.

In my tests for compression the standard -6 on PIGZ is beaten by zstd for ratio by zstd lvl3 and completes 443 seconds faster… We could up to zsd lvl11 and have it 93 seconds quicker and save around 550MB.

[Junkhacker]

i’ve been talking with @Tom-Elliott about this, and we don’t think it would be worth the effort it would take to implement zstandard. the thing is, faster decompression is kind of irrelevant for FOG at the moment. what slows down deployments at the moment is transfer speed. the only way fog would get faster is if the file size was very significantly decreased. while the compression ratio is a better with zstandard, the difference isn’t very significant until you get to the higher compression levels, where processing time becomes a big issue.

there are other issues that deter us from adoption, but that’s the most significant reason. in fact, the single greatest reason TO adopt it would be because i think it’s really cool, lol.

[Jaymes Driver]

@Junkhacker said in ZSTD Compression:

there are other issues that deter us from adoption, but that’s the most significant reason. in fact, the single greatest reason TO adopt it would be because i think it’s really cool, lol.

It would be REALLY COOL

[Tom Elliott]

THe problem isn’t the implementation or not.

Already, with PIGZ in use the issue (beyond multiple Unicast tasks) is most often slow down in writing the information to the disk. This is especially present when one is dealing with SSD.

It’s great that you can have “fast” decompression, but that only goes so far. You still have to write the data to disk. You have some buffer, but we’re already “decompressing” the data as fast as we can.

Where this might be very useful, however, would be uncompressed images, compressed as the data is requested, and then placed on disk so we have a live element of diminishing the amount of data to be passed across the network. Once it’s passed to the client, the only “hold” is on the speed at which data can be pushed from ram and written to disk. Even this, however, can only do so much.

Is it really worth implementing a new compression mechanism to maybe get a speed increase of possibly 1% during our imaging process?

[Tom Elliott]

I understand the speed would be significantly increased on upload tasks, but I don’t know how often people are uploading.

[VincentJ]

1 vCPU 1.6GHz - the system can no longer saturate gigabit over network shares…
Down from 110MB/s to 82MB/s

Compression - Compressed size - Decompression time
zstd lvl1 - 7,940,779KB - 131 seconds
zstd lvl3 - 7,420,268KB - 134 seconds
zstd lvl11 - 6,967,155KB - 139 seconds
zstd lvl22 - 6,214,702KB - 157 seconds

pigz.exe --keep -6 a:\d1p2 - 7,535,149KB - 247 seconds

On my quad core VM PIGZ -6 only used 50MB/s decompression, zstd level 11 with a single core VM uses the same 50MB/s…
On the single core VM, PIGZ -6 is only 30 MB/s, the lowest zstd gets on level 22 is 39.5MB/s

if we use the single core numbers, writing the whole image in 247 seconds (which isn’t too much faster than expected anyway) is around 66MB/s on disk, using zstd 11 writing it in 139 seconds is 117MB/s Most SATA disks should be able to do this… It will be a push for some 2.5" disks… (I checked numbers for 2.5" and 3.5" WD Greens)

[compman]

Note that, since v1.1.3, there is a multithread mode available with zstd.

It needs to be compiled with specific flags though.
On linux, it means typing make zstdmt
For Windows, there are pre-compiled binaries in the release section : use the zstdmt one.

Since pigz is multi-threaded, it would be more fair to compare to zstdmt, rather than single-threaded zstd.

The number of threads can be selected with command -T#, like xz.

[VincentJ]

The version of zstd i’ve been using is using all my threads

[VincentJ]

Maybe you just saw the note about 1vCPU. I only reduced to 1vCPU as the numbers with 4vCPU were all so close together.

Also might help to simulate a ‘low end’ machine…

[A Former User]

For those of us not smart enough to fully understand, can someone give me a simple comparison – in time – between the proposed compression versus the current compression?

I work two IT jobs, one full-time and one part-time. Between the two, we order tens of thousands of computers each year. For January 2017, this was the most ordered machine from Dell, and most other orders were also in this same power range:

OptiPlex 5040 Small Form Factor
i5-6500 Processor (Quad Core, 6MB, 3.2GHz)
8 GB RAM
256 GB SSD
1 Gbps NIC

Percentage wise, approximately how much faster/slower would the proposed compression be for this machine when deploying an image to it?

[Tom Elliott]

@loosus456 it’s hard to say. Compression on upload would be phenomenal buy for deployment I don’t think there’d be a huge difference as even with our current stuff we’re mostly seeing speeds to write to disk.

[A Former User]

@Tom-Elliott We do upload often (about twice a month), but if the upload isn’t much, much faster and the deployment isn’t significantly faster, it probably isn’t worth it.

I do wonder if HyperV upload through the legacy adapter would be faster, though. That takes literal hours right now.

[Tom Elliott]

@loosus456 Let’s say you upload 2 image’s a month, and you deploy 400 times a month, ultimately while upload would be “faster” you’re only increasing it during the upload process. As you still have your “setup” to create the image which is what’s taking the most of your time.

[A Former User]

@Tom-Elliott Well, when it comes to uploading from HyperV with a legacy adapter to FOG, upload time is actually what takes most of the time. Image creation takes little time in comparison.

But yes, uploading from a physical machine is quite fast.