Zswap Could See Better Performance Thanks To A B-Tree Search Implementation

[phoronix]

Phoronix: Zswap Could See Better Performance Thanks To A B-Tree Search Implementation

For those using Zswap as a compressed RAM cache for swapping on Linux systems, the performance could soon see a measurable improvement...

http://www.phoronix.com/scan.php?page=news_item&px=Zswap-Linux-B-Tree-Search

[S.Pam]

No, imagine if they would also allow swapping out compressed pages. Currently they decompress before saving to disk!

[ermo]

No, imagine if they would also allow swapping out compressed pages. Currently they decompress before saving to disk!

Wait, what?!

I wonder why they don't just keep the pages compressed as that might a) Make for smaller writes and b) Make for faster reads and c) Result in more effective space for swap data?

When using z3fold and lz4, I would imagine that most modern CPUs can keep up with reading compressed pages from disk and decompressing them on the fly almost transparently...?

[S.Pam]

Indeed. The problem seems to be that they would need a new on-disk format or logic to separate compressed vs uncompressed pages on disk. So instead of doing that, they decompress before writing to disk. This is a huge overhead.

[alpha_one_x86]

Integration to kernel 5.5 planned? or 5.4?

[pomac]

wrt writing down the compressed data, they would still write 4k entries - and the decompression is just about no overhead... I wonder how much you'd gain.

[Veerappan]

Regarding the performance: Yes please!

I've been using zswap since I got my current work laptop (soon to be replaced, hopefully). Compressed pre-swap is the only reason that this machine (T440p, quad-i7 haswell, 16GB RAM, 500GB SSD) was usable in the first place for what I do. There's still a noticeable impact to the machine's performance once zswap kicks in, and hopefully this makes that more tolerable.

Of course, a new laptop with at least 32GB should hopefully be right around the corner... As long as it's got at least 8 threads to go with it and isn't crippled long-term by nvidia graphics.

[betam4x]

Regarding the performance: Yes please!

I've been using zswap since I got my current work laptop (soon to be replaced, hopefully). Compressed pre-swap is the only reason that this machine (T440p, quad-i7 haswell, 16GB RAM, 500GB SSD) was usable in the first place for what I do. There's still a noticeable impact to the machine's performance once zswap kicks in, and hopefully this makes that more tolerable.

Of course, a new laptop with at least 32GB should hopefully be right around the corner... As long as it's got at least 8 threads to go with it and isn't crippled long-term by nvidia graphics.

I run on a 32 GB system with no swap, quite simply because when I had a swap partition, it constantly showed 0% utilization. My workloads always used around 20GB tops of RAM and for whatever reason the kernel didn't feel the need to page seldom used pages to disk. That being said, on a desktop at least, I've had few issues with NVIDIA drivers outside of Wayland support under KDE. However, as Wayland itself is immature, I really don't care all that much currently, and NVIDIA is working on a permanent fix for Wayland that will eventually make it out. While I'm definitely not for a monopoly, NVIDIA is applying much needed competitive pressure to AMD to get them to step up to the plate performance wise. If AMD hadn't been making missteps since the Radeon 3xx series, they wouldn't be in the same boat they are in now. I'm not going to reward that incompetence. Yeah NVIDIA only has a blob driver that performs well, but most people don't look at the source code anyway. I see things as more of an ebb and flow. If AMD starts beating NVIDIA consistently again, NVIDIA will have to differentiate themselves and improve. This will probably mean opensourcing their driver stack at some point, especially if Intel storms onto the scene with open source drivers and a strong GPU offering.

[pal666]

If AMD starts beating NVIDIA consistently again, NVIDIA will have to differentiate themselves and improve. This will probably mean opensourcing their driver stack at some point, especially if Intel storms onto the scene with open source drivers and a strong GPU offering.

so nvidiots are just going to sit and wait for nvidia to maybe throw them bone

[finalzone]

That being said, on a desktop at least, I've had few issues with NVIDIA drivers outside of Wayland support under KDE. However, as Wayland itself is immature, I really don't care all that much currently, and NVIDIA is working on a permanent fix for Wayland that will eventually make it out.

It seems the KDE implementation of Wayland is immature in that context. Nvidia could simply use the existing implementation to enable such support for Wayland protocol like both AMD and Intel did.

While I'm definitely not for a monopoly, NVIDIA is applying much needed competitive pressure to AMD to get them to step up to the plate performance wise.

The pressure is on Nvidia this time as AMD open source driver (RadeonSI and RADV) is gradually closing the gap and the fact majority of distribution users can play heavy 3D games out of box with the latter. Both Valve and Google stepped in the game to improve the performance of open source driver. Nvidia secret is convincing developers to use their non-standard method in detriment of competitors. Majority of PC gaming developers have no one than themselves with the current condition of PC games.