Nvidia says its SmartNICs sizzled to world record storage schlepping status Nvidia has claimed a world record for storage IOPS using its Bluefield data processing units (DPUs). The Bluefield kit is Nvidia’s take on a “SmartNIC” – a network interface card bulked up with a decent CPU so it can run workloads such as firewalls or encryption engines, so that CPUs can be freed for more important tasks. …
... in cases where the standby system and database instance(s) are physically remote from the live instance.
Typically this type of replication means sending copies of the live system's transaction log to the backup system(s) and applying the transferred log items to the standby database(s) in as close to real-time as possible. Speeding up transaction log transfer minimizes cut-over time because a standby system must not accept the live workload until its database is fully up to date.
I'm pretty sure people still remember the Killer NIC from Rivet Networks. Everyone I know was calling it snake oil that Killer quickly went bankrupt and was bought by Atheros who only used the brand to make run-of-the-mill cards with existing chips, only with custom firmware and drivers, and then resold the brand to Intel lately. Surely this is no different?
I think in principle it's exactly the kind of thing Rivet Networks did just at a different scale. I remember the Killer NIC, it was marketed to gamers; it would bypass Windows' TCP/IP stack in favor of handling that on the card, plus it had like FTP and bittorrent apps that could run on the card. But it was pricey, and the NIC was gigabit, but if you had any normal internet connection (this was 15 years ago), how much CPU time does it take to handle like 10mbps anyway?
Now, there's two factors on this current stuff. 1) These hyperscalers are going for the "mainframe model", leaving the CPUs free to compute with IO processors handling disk I/O, network I/O, etc. as much as possible. If you're running 100gbps the traditional way that'll keep several CPUs very busy just handling interrupts and whatever. 2) Running some traditional router functionality on there, they may be able to buy very expensive NICs in lieu of even more expensive switches and routers.
The wheel turns..
30 years ago, the processing for things like Modems, Network cards and even laser printers was handled on the device. Then, to reduce the number of parts, and therefore, the cost, they moved most of the processing for these devices to the computer's CPU, leaving the device fairly dumb.
The same applies to some extent to Audio and Video, but for a long time, users in those spaces have had the option of using the onboard chipsets for low end stuff, and cards or even external adaptors (using USB, Thunderbolt or whatever) for high end needs.
Of course, for low end uses such as printing, using a modem etc, the devices didn't put much of a load on the CPU, even NICs with 10 or 100Mbs networking.
Now, we come full circle, with devices doing as much of their own processing as possible, to reduce the load on the CPU(s).
Well, not even 20 years ago the green team was offering chipset for the red team that implemented some form of network firewall. The execution was imperfect and they discontinued software support for this product before pulling out completely from the x86 chipset business. I'd like to know if any of that legacy has lived on.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
