Btrfs Allocator Hints

Sun's ZFS7420 had a terabyte of RAM per controller, and these work in tandem, and after a certain pressure, the thing can't keep up even though it also uses specialized SSDs to reduce HDD array access during requests, and these were blazingly fast boxes for their time.

When you drive a couple thousand physical nodes with a some-petabytes sized volumes, no amount of RAM can save you. This is why Lustre divides metadata servers and volumes from file ones. You can keep very small files in metadata area (a-la Apple's 0-sized, data-in-resource-fork implementation), but for bigger data, you need to have good filesystems. There are no workarounds from this.

If you want to go faster, take a look at Weka and GPUDirect. Again, when you are pumping tons of data to your GPUs to keep them training/inferring, no amount of RAM can keep that data (or sustain the throughput) during that chaotic access for you.

When we talked about performance, we used to say GB/sec. Now a single SSD provides that IOPS and throughput provided by storage clusters. Instead, we talk about TB/sec in some cases. You can casually connect terabit Ethernet (or Infiniband if you prefer that) to a server with a couple of cables.

ZFS can, and is actually designed to, handle that kind of workloads, though. At full configuration, ZFS7420 is a 84U configuration. Every disk box has its own set of "log" SSDs and 10 additional HDDs. Plus it was one of the rare systems which supported Infiniband access natively, and was able to saturate all of its Infiniband links under immense load.

Lustre's performance is not RAM bound when driving that kind of loads, this is why MDT arrays are smaller and generally full-flash while OSTs can be selected from a mix of technologies. As I said, when driving that number of clients from a relatively small number of servers, it's not possible to keep all the metadata and query it from the RAM. Yes, Lustre recommends high RAM and core count for servers driving OSTs, but it's for file content throughput when many clients are requesting files, and we're discussing file metadata access primarily.

These same people are the only ones in the world suggesting btrfs is "basically" stable. I'll never touch this project again with a ten foot pole, afaic it's run by children. I'll trust adults with my data.

Moreover, you again give examples from the end product. Finished sites, compacted JS files, compressed videos, compiled models...

There's much more than that. The model is in RAM, but you need to rake tons of data over that GPU. Sometimes terabytes of data. You have raw images to process, raw video to color-grade, unfiltered scientific data to sift through. These files are huge.

A well processed JPG from my camera is around 5MB, but RAW version I process is 25MB per frame, and it's a 24MP image, puny for today's standards. Your run of the mill 2K video takes a couple of GBs after final render at movie length. RAWs take 10s of terabytes, at minimum. Unfiltered scientific data again comes in terabytes to petabytes range depending on your project and instruments you work on, and multiple such groups pull their own big datasets to process real-time.

In my world, nothing fits in RAM except the runtime data, and that's your application plus some intermediate data structures. The rest is read from small to gigantic files and written in files of unknown sizes, by multiple groups, simultaneously. These systems experience the real meaning of "saturation", and they would really swear at us at some cases.

Sorry, but you can't solve this problem by buying more RAM, because these workloads can't be carried to clouds. They need to be local, transparent and fast. IOW, you need disk systems which feel like RAM. Again, look what Weka (https://www.weka.io/) does. It's one of the most visible companies which make systems behave like a huge RAM, but with multiple machines and tons of cutting edge SSDs, because what they process doesn't fit in RAM.

Lastly, oh, there's a law which I forget its name every time, which tells you if you cache 10 most used files, you can serve up to 90% of your requests from that cache, if your request pattern is static. In cases I cite, there's no "popular" file. Everybody wants their own popular files which makes access "truly random".