The Opening Keynote at SREcon APAC

Brendan Gregg delivered the opening keynote at this year's USENIX SREcon APAC, covering the current state and likely trajectory of computer performance. The material draws from updates made for the second edition of Systems Performance, and this marked the first time the talk was given in person. The full session is available on YouTube, with the slide deck posted both online and as a PDF for those who want to follow along at their own pace.

The CXL Question

During the Q&A portion, Gregg was asked about Compute Express Link (CXL), which he admitted was a lucky break—he had intended to discuss it but had forgotten to include it in the main deck. CXL effectively enables a custom memory controller to be attached to a system, expanding memory capacity, bandwidth, and overall performance. Gregg's personal assessment is cautious: he doesn't see widespread demand for more capacity given the prevalence of horizontal scaling and servers that already support over 1 Tbyte of DRAM. Additional bandwidth would be useful, but the extra hop required to reach that memory introduces latency that gives him pause. In short, it's an interesting technology, but the killer use case has not yet emerged.

From Junior Sysadmin to Keynote

Gregg's path to the SREcon stage began in 1999 as a junior Unix sysadmin in Newcastle, New South Wales, far from the tech hubs of Silicon Valley or even nearby Sydney. Determined to advance regardless of geography, he consumed every sysadmin book and magazine he could find, including SysAdmin magazine. Two particular items stood out: a call for article submissions handled by editor Rikki Endsley, and an advertisement for USENIX conferences where one could learn directly from industry experts. Despite never having been published, traveled to the US, or even flown, Gregg set both as goals.

While he never published directly in SysAdmin, his performance work later ran as a feature article. His USENIX involvement began in 2010 when a community manager encouraged him to start attending and speaking. Since then, he has built lasting connections-including with Amy Rich, now USENIX President, and Rikki Endsley, with whom he co-chaired USENIX LISA18. Gregg credits USENIX as a significant help to his career and his employers, and emphasizes the value of the organization's vendor-neutral space for sharing technical advances.

The chance to help bring USENIX conferences to Australia with the opening keynote was, in his words, something he could never have scripted or expected.

Next Stop: SREcon 2023 APAC

With the 2022 conference behind him, Gregg is already looking ahead. He is serving as program co-chair for SREcon 2023 APAC, scheduled for June 14-16 in Singapore. The call for participation closes on March 2nd at 23:59 SGT—roughly 24 hours from the time of this writing.

Opening Notes from SREcon APAC

Brendan Gregg opened USENIX SREcon APAC 2022 with a keynote on where computer performance is headed, drawing on material from the second edition of Systems Performance and giving the talk in person for the first time. The video is posted on YouTube, with the slides online and as a PDF.

CXL: Still Hunting for Its Killer App

The Q&A portion of the talk covered Compute Express Link, a topic Gregg had planned to include but forgot to fold into the main presentation. The question was welcome, as it gave him a chance to address it directly. CXL essentially allows a custom memory controller to be attached to a system, expanding memory capacity, bandwidth, and overall performance. Gregg's view is measured: he does not see a widespread need for additional capacity given current horizontal scaling and servers that can already exceed 1 Tbyte of DRAM. Extra bandwidth is helpful in some cases, but he worries about the latency hit from adding a hop to reach the additional memory. The technology is interesting, he said, but the compelling use case is not there yet.

A Long Road to the Keynote Stage

Gregg began his career in 1999 as a junior Unix sysadmin in Newcastle, NSW, with no ties to the tech scenes in Silicon Valley, New York, or even nearby Sydney. Determined to master his new field from that distance, he read every relevant book, article, and magazine he could find. That included SysAdmin magazine, where two details stood out: an invitation for readers to submit their own articles, and an advertisement for USENIX conferences where one could learn directly from the experts. Both became explicit goals, even though he had never been published, never visited the US, and never even boarded a plane.

His work eventually appeared as a feature article in that magazine, and he began attending and speaking at USENIX conferences in 2010 after encouragement from a community manager. Over time he built lasting connections, including with Amy Rich, now USENIX President, and Rikki Endsley, with whom he co-chaired LISA18. He credits USENIX with being a major benefit to his career and employers, and calls it an important vendor-neutral venue for exchanging technical knowledge. Bringing USENIX to Australia as an opening keynote presenter was, in his words, something he could not have scripted or expected.

Looking to SREcon 2023

With little time to dwell on the past, Gregg has moved on to planning the next event. He is program co-chair for SREcon 2023 APAC, taking place June 14-16 in Singapore. The call for participation closes March 2nd at 23:59 SGT—roughly 24 hours from now.

A Note on the SREcon APAC 2022 References

The source article from Brendan Gregg includes the full reference list for his USENIX SREcon APAC 2022 keynote. He makes a deliberate editorial choice here about citation style: he includes author names alongside talk titles and dates for every entry, rather than the more common practice of simply listing bare URLs.

This attention to attribution is something Gregg notes he adopted after realizing how "unfair" it was that some references in his earlier books had named authors while others did not. His rule now, he writes, is to always include full names when they are known.

The list itself is a comprehensive survey of the performance landscape his talk covers, spanning:

  • Memory and storage: Entries cover everything from DDR4 and DDR5 latency characteristics to HBM introduction, Intel Optane’s rise and fall, and the evolving HDD technologies such as SMR and HAMR from Seagate and Western Digital.
  • Networking and offload: The list includes foundational congestion control algorithms like DCTCP and BBR, multipath TCP extensions, and hardware offload work like FreeBSD’s kernel TLS. It also tracks the movement of packet processing into hardware, including FFShark’s FPGA implementation of BPF filtering at 100G.
  • Silicon and packaging: The references chronologically trace the shift toward heterogeneous and tiled topologies, including Intel’s EMIB, AMD’s 3D V-Cache, and Sapphire Rapids’ modular SoC design, alongside the broader process node roadmap down to 2nm.
  • System software and observability: The bibliography includes the major milestones in Linux performance tooling from BPF superpowers to flame graphs, along with newer kernels that brought in features like AutoFDO. It also points to case studies of in-kernel acceleration, such as BMC for Memcached.
  • Compute specialization: The list reflects the widening ecosystem of custom and dedicated processors, from Google’s TPU and Azure’s Ampere Altra instances to the massive parallel designs from Cerebras and Esperanto, as well as the rise of SmartNICs and infrastructure processing units (IPUs).

From a professional editor's perspective, the bibliography is notable not just for its range — from leading academic papers to news coverage and vendor pages — but for the way it grounds an optimistic roadmap in concrete references that practitioners can follow up on themselves. It is a valuable reading list for anyone trying to understand where the performance gains of the last decade and the next one have come from.