Qumulo debuts machine learning-driven caching for faster data access

Improvements in Qumulo’s QF2 software means its products run faster –  20 per cent quicker than Dell EMC Isilon boxes, by one measure.


Qumulo has juiced up QF2 with what it calls machine learning-powered pre-fetch – in other words, the software queues up the files that  it “thinks” users will most likely access next.  QF2 can supposedly recognise user/workload specific behaviour and carries out dynamic workload scaling to maximise cache use.

It performs file access pattern analysis and proactively caches individual file blocks that are most likely to be repeatedly accessed.

Qumulo says the software-driven speed increase mean the all-flash P-Series can do 6K video playback and the hybrid QC_Series uncompressed 4K video playback. Users could see a 3x-10x performance improvement in frame-based workloads across video editorial and machine learning.

Blocks and Files’s hunch is that other filer software vendors will copy this set of caching improvements.

Qumulo first announced all-flash systems in November last year, in the shape of the P-Series

The hardware was a 2U Intel box that can run up to 24 NVMe SSDs in U.2 format, dual Xeon Skylake Gold-level 6126 processors and 192GB (12 x 16GB) of RAM. There were two models; the 23T is a 23TB model with 12 x 1.92TB SSDs while the 92T has a 92TB capacity using 24 x 3.84TB SSDs. The drives use 3bits/cell (TLC) 3D NAND.

The 92T ten pumped out 16GB/sec; 4GB/sec from a single node. Now the 92T offers 18GB/sec throughput and 200,000 IOPS from a four-node cluster. Qumulo says this means 90GB/sec from a rack.

In comparison Dell EMC’s all-flash Isilon F800 filers offer up to 15GB/sec per chassis and up to 250,000 IOPS; ahead on the IOPS front but behind on the throughput measure. But then it does have four Xeon CPUs per chassis instead of Qumulo’s two.