Memory after DRAM

I spent nearly a decade at Samsung, one of the world’s largest DRAM vendors, and now I work on replacing DRAM. The devices keep improving, but progress in cost per bit, latency, and energy efficiency has slowed. Alternative memories have been “emerging” since 2001, and several are now good enough to substantially reduce how much DRAM a real system needs. I enjoy making that practical: redesigning hardware on an FPGA, changing the Linux kernel to make placement decisions cheap and effective, and rethinking the interface between the two to minimize overhead.

I am advised by Philip Levis in the Stanford Information Networks Group, and much of this work is part of Stanford Differentiated Access Memory (DAM).

Research

Application read-only memory

Denser and cheaper memories bring constraints DRAM does not have: asymmetric read and write cost, limited endurance, and coarse write granularity. Commercial parts hid those behind a translation layer on the module, and the layer cost more than it saved. I work on the opposite arrangement, where the system exposes the constraints and the kernel manages them. Such memory is read-only to applications and written only by the kernel during migration, enforced with copy-on-write, which moves wear levelling off the device and lets the hardware get simpler rather than more complex.

Application DRAM small, writable Long-term RAM large, cheap, finite writes read and write read only kernel migrates
Paper
  • The Kernel's Write: Application Read-Only Memory
    Hui Sub Shim, Katherine Mohr, Philip Levis
    4th Workshop on Disruptive Memory Systems (DIMES '26), 2026

Packet processing and memory bandwidth

Network interfaces now have enough compute to inspect traffic at line rate and nowhere near enough memory bandwidth to feed it. Packets arrive variable sized and out of order while the memory they land in is a contiguous address space divided into fixed-size pages, and translating between the two layouts costs copies that consume the bandwidth the analysis needed. With colleagues I work on addressing that lets software read a whole flow as one contiguous region, built and measured on FPGAs. On an Alveo U250 prototype it cuts DRAM traffic by 77 to 83%, lowers execution time by 25 to 62%, and raises per-core throughput by 1.3 to 2.7 times.

packets arrive variable sized, out of order 3 1 4 2 memory is fixed-size pages copies, every byte instead one contiguous flow no copies
Paper
  • Packets are Not Pages: Flow-Based Addressing Conserves Memory Bandwidth
    Agur Adams, Hui Sub Shim, Colin Drewes, Ben O'Keefe, Philip Levis, Zakir Durumeric
    Proceedings of the ACM on Networking, Vol. 4, CoNEXT3 (CoNEXT '26), 2026

Placing data across unlike memories

Once a server holds several kinds of memory, the operating system has to decide what belongs where, and how often a page is touched turns out to be too coarse a signal. Two pages read equally often can perform very differently in the same memory, because memory-level parallelism, compression and read and write asymmetry all cut across access frequency. I am interested in what the hardware should tell the operating system about how well a page suits a given memory, and in measuring what deployed disaggregated memory actually costs.

two pages, touched equally often page A page B 100 reads/s 100 reads/s fits: parallel, compressible does not fit: serialised fast slow how often a page is touched does not say which memory suits it