I’ve just started using zram for swap on VMs. The use of compression for swap in Linux apparently isn’t new, it’s been in the Linux kernel since version 3.2 (since 2012). But until recent years I hadn’t used it. When I started using Mobian (the Debian distribution for phones) zram was in the default setup, it basically works and I never needed to bother with it which is exactly what you want from such a technology. After seeing it’s benefits in Mobian I started using it on my laptops where it worked well.
Benefits of ZRAM
ZRAM means that instead of paging data to storage it is compressed to another part of RAM. That means no access to storage which is a significant benefit if storage is slow (typical for phones) or if storage wearing out is a problem.
For servers you typically have SSDs that are fast and last for significant write volumes, for example the 120G SSDs referenced in my blog post about swap (not) breaking SSD [1] are running well in my parents’ PC because they outlasted all the other hardware connected to them and 120G isn’t usable for anything more demanding than my parents use nowadays. Those are Intel 120G 2.5″ DC grade SATA SSDs. For most servers ZRAM isn’t a good choice as you can just keep doing IO on the SSDs for years.
A server that runs multiple VMs is a special case because you want to isolate the VMs from each other. Support for quotas for storage IO in Linux isn’t easy to configure while limiting the number of CPU cores is very easy. If a system or VM using ZRAM for swap starts paging excessively the bottleneck will be CPU, this probably isn’t going to be great on a phone with a slow CPU but on a server class CPU it will be less of a limit. Whether compression is slower or faster than SSD is a complex issue but it will definitely be just a limit for that VM. When I setup a VM server I want to have some confidence that a DoS attack or configuration error on one VM isn’t going to destroy the performance of other VMs. If the VM server has 4 cores (the smallest VM server I run) and no VM has more than 2 cores then I know that the system can still run adequately if half the CPU performance is being wasted.
Some servers I run have storage limits that make saving the disk space for swap useful. For servers I run in Hetzner (currently only one server but I have run up to 6 at various times in the past) the storage is often limited, Hetzner seems to typically have storage that is 8* the size of RAM so if you have many VMs configured with the swap that they might need in the expectation that usually at most one of them will be actually swapping then it can make a real difference to usable storage. 5% of storage used for swap files isn’t uncommon or unreasonable.
Big Servers
I am still considering the implications of zram on larger systems. If I have a ML server with 512G of RAM would it make sense to use it? It seems plausible that a system might need 550G of RAM and zram could make the difference between jobs being killed with OOM and the jobs just completing. The CPU overhead of compression shouldn’t be an issue as when you have dozens of cores in the system having one or two used for compression is no big deal. If a system is doing strictly ML work there will be a lot of data that can’t be compressed, so the question is how much of the memory is raw input data and the weights used for calculations and how much is arrays with zeros and other things that are easy to compress.
With a big server nothing less than 32G of swap will make much difference to the way things work and if you have 32G of data being actively paged then the fastest NVMe devices probably won’t be enough to give usable performance. As zram uses one “stream” per CPU code if you have 44 cores that means 44 compression streams which should handle greater throughput. I’ll write another blog post if I get a chance to test this.
Leave a Reply
You must be logged in to post a comment.