Summary: <br> <br> <br> <br> <br> This week’s podcast episode continues the discussion on SmartNICs and DPUs with <a href="https://www.fungible.com/">Fungible</a>, a company that claims to have originally coined the term DPU. Chris and Martin talk with Pradeep Sindhu (CEO and co-founder) and Jai Menon (Chief Scientist) about Fungible’s storage cluster and host-based DPU. The Fungible architecture aims to solve the challenges of disaggregation, a topic we first looked at back in <a href="https://storageunpacked.com/2017/09/garbage-collection-005-disaggregated-storage/">September 2017</a>. <br> <br> <br> <br> This discussion highlights some interesting challenges that new technology such as NVMe-oF is introducing into the data centre. As we move to a model of highly parallelised workloads, the interaction between storage and compute is back under the spotlight. Fungible is working on storage products today but claims to be able to disaggregate GPUs and in the future, potentially DRAM. Interesting times.<br> <br> <br> <br> Find out more at <a href="https://www.fungible.com/">https://www.fungible.com/</a><br> <br> <br> <br> Elapsed Time: 00:49.25<br> <br> <br> <br> Timeline<br> <br> <br> <br> * 00:00:00 – Intros* 00:01:30 – What problem is Fungible looking to solve?* 00:02:45 – The benefits of Moore’s Law growth are almost flat* 00:03:35 – Modern applications are data-centric* 00:05:00 – Fungible is working on disaggregated architectures* 00:06:15 – There’s storage for compute and storage for – storage!* 00:07:45 – What’s “hyper-disaggregation”?* 00:09:30 – Fungible offers volume-specific characteristics like encryption* 00:11:00 – Everything can be disaggregated except DRAM (for now)* 00:13:10 – Data I/O has specific requirements including in-order processing* 00:14:15 – Fungible can use an Ethernet network at 90% without packet drop* 00:15:30 – Data-centric workloads are heavily multiplexed* 00:18:45 – Does disaggregation finally deliver a real software-defined data centre?* 00:20:00 – Mainframes “reconfigured” overnight for batch workloads* 00:21:30 – Even hyperscalers operate in silos* 00:24:15 – The Fungible Storage Cluster – SAN 2.0?* 00:25:30 – New hero numbers! 15 million IOPS!!* 00:30:45 – Volumes are virtual across any or all storage clusters* 00:34:30 – Fungible claims better than local performance* 00:35:00 – The Storage Cluster gains additional benefits with a host DPU* 00:37:00 – Let’s not get diverted towards VSAM!* 00:39:20 – Disaggregated technologies could deliver truly reconfigurable data centres* 00:41:00 – Fibre Channel networks divide their traffic across multiple SANs* 00:42:00 – Who is using the Fungible technology?* 00:45:00 – Fungible is working on I/O primitives for SQL databases* 00:49:00 – Wrap Up<br> <br> <br> <br> Related Podcasts & Blogs<br> <br> <br> <br> * <a href="https://storageunpacked.com/2021/02/194-scaleflux-csd/">#194 – ScaleFlux & Computational Storage Devices</a>* <a href="https://storageunpacked.com/2021/01/190-nvidia-smartnics-dpus/">#190 – NVIDIA BlueField SmartNICs & DPUs</a>* <a href="https://storageunpacked.com/2020/10/180-pliops-storage-processor/">#180 – SmartNICs – Pliops Storage Processor</a>* <a href="https://storageunpacked.com/2020/10/177-smartnics-project-monterey/">#177 – SmartNICs and Project Monterey</a>* <a href="https://storageunpacked.com/2019/05/96-smartnics-mellanox-rob-davis/">#96 – Discussing SmartNICs and Storage with Rob Davis from Mellanox</a><br> <br> <br> <br> <br> <br> <br> <br> Pradeep’s Bio<br> <br> <br> <br> <br> <br> <br> <br> Pradeep’s career includes founding Juniper Networks,
