Alibaba Cloud details storage tech that's doubled its VMs per host

"Exclusive Alibaba Cloud has detailed the tech it developed to run local storage in its servers and bust bottlenecks created by new-generation manycore processors.

The tech was detailed last week in a paper titled ""CSAL: the Next-Gen Local Disks for the Cloud"" published in the April edition of Proceedings of the Nineteenth European Conference on Computer Systems. Eleven authors work at Alibaba Cloud, and another six work at Solidigm – Intel's old SSD business now mostly owned by SK hynix.

The paper sets the scene by reminding readers that cloud servers typically use local storage, and that local capacity determines how many VMs each cloudy host can handle. It then notes that modern manycore CPUs encourage clouds and users to run more VMs on each host.

The obvious way to run more VMs, the paper notes, is to pack cloud servers full of colossal hard disks, storage-class memory, or fast solid state disks. But hard disks have bandwidth limits, storage-class memory mostly failed (the paper mentions the Optane tech Intel snuffed), and fast SSDs have capacity problems and big price tags.

What's a cloud to do? Quad-level cell (QLC) SSDs are an obvious answer, the paper suggests, because they offer high capacity and decent prices.

Alibaba Cloud therefore tried QLC disk in three scenarios: as a drop in replacement for other disks, as part of a layered system alongside high-speed SSDs, and using the dm-zoned a kernel device mapper.

A layered system that used a write-back cache to handle small writes in one SSD helped, but didn't match hard disk performance.

dm-zoned didn't help either, because under load it constantly needed to move data – which smashed performance.

Alibaba therefore devised the Cloud Storage Acceleration Layer (CSAL), which the paper explains sees the most recently used data stored in DRAM and swapped to a fast SSD, which also handles all incoming writes. When possible and sensible, data from that SSD is shunted into the QLC disk.

The paper explains CSAL's workings in sufficient detail that that even our storage-centric sibling site Blocks and Files might find its attention wavering.

The impact of CSAL on Alibaba Cloud ops is easier to understand and is outlined as follows:

Compared to last-gen HDD-based local disks (24× 2TB HDDs with a 48-core Xeon Cascade CPU), CSAL-ready servers (an 800GB HP-SSD and a 15.36TB QLC SSD with a 64-core Xeon Ice Lake CPU) can host twice more instances while achieving the same Service Level Objects.

That's a doubling of VM density from second-gen to third-gen Xeons, despite the extra 16 cores in the newer processor stressing storage more than the older silicon. Also, 64 is not double 48.

CSAL is in production across ""thousands of Elastic Compute Service (ECS) nodes in Alibaba Cloud."" Maybe you could run it too: Alibaba Cloud has open-sourced CSAL into the Storage Performance Development Kit.

Alibaba Cloud has racked up a few wins lately. After cutting prices, its homebrew Yitian 710 was recently rated the fastest Arm CPU in the cloud. We've also covered an in-house networking tool that slashed the number of personnel the Chinese concern needed to dedicate to troubleshooting, and research suggesting Alibaba Cloud's operations could be more efficient than Google's."

Author

ABN ASIA was founded by people with deep roots in academia, with work experience in the US, Holland, Hungary, Japan, South Korea, Singapore, and Vietnam. ABN Asia is where academy and technology meet opportunity. With our cutting-edge solutions and competent software development services, we're helping businesses level up and take on the global scene. Our commitment: Faster. Better. More reliable. In most cases: Cheaper as well.

Feel free to reach out to us whenever you require IT services, digital consulting, off-the-shelf software solutions, or if you'd like to send us requests for proposals (RFPs). You can contact us at contact@abnasia.org. We're ready to assist you with all your technology needs.

© ABN ASIA