ASPLOS '06

From PEL Wiki

This is a transcription of my notes from ASPLOS '06 in San Jose. Maybe I'll try to clean them up later.

VMWare Keynote

Application specific OS? Aren't there enough of them already?

At home each child gets a virtual PC.

Other virtualization stuff

Kelly and I talked about stripping down the OS automatically for distribution in a VM for servers with the application. There's no need to support all the bells and whistles of an OS for each App. Could you even strip out protection to get back Jim Larus' 37%?

Wild and Crazy Ideas Session

Jim Larus - Microsoft

37% overhead for not using type-safe languages. This is in terms of operating system overhead and machine overhead for switching user-levels and for page protections.

Ivan Sutherland - Berkeley

Fleet - one instruction computer. No PC, no jumps, no code blocks, and no sequential programming model.

Historically misplaced jumps. At the beginning of the code segment would be a lot more useful for speculation, etc.

Chip Recycling

Re-use chips (yeah ... right) from high end to low end.

Interesting random fact: it takes the energy equivalent of about 200 gallons of gasoline to produce one 300mm wafer.

Colorama

Use colors for data-centric model of critical sections. It sounds good, but what do you do when there are shared ones?

Cell phone clusters

John Shen at Nokia Research Center said "Develop the course for distributed computation with 100 cell phones." It sounded like he'd supply the phones.

Transactional Memory

VMWare could implement a transactional memory VM like IBM did in the AS/400 and System/38.

One of the papers decided to always commit the transactions in place but keep a shadow page for roll-backs. It seemed like it made more sense to update the shadow and then commit if there wasn't an abort. That way you did fewer copies every time.

Their way

1. Copy to shadow page
2. Update in place
3. if (abort) Copy from shadow

The other way

1. Update shadow page
2. if (!abort) Copy from shadow

If abort never happens they are equal (1 copy, 1 update, 1 test) If abort happens at all, the other way wins.

Stephen (Todd) Jones' paper

He showed a 3 second delay between when the OS kicked out a page from the disk cache and when his system found out about it. I was surprised it was so uniform. It seemed interesting to note for DiskRAM.

Memory Debugging

The Heap server paper has #s for memory allocations per second for several benchmarks. They do bug finding based on heap stats.

Geiger

DiskRAM filesystem pitfalls? No journalling? Submit IEEE S&P 20006

Memory Virtualization

This is something I need to understand better. AMD's version is called Nested Page Table Intel's version is called Extended Page Table

Scatter gather buffers in memory to avoid moving memory around?

People

John Shen has left Intel and joined Nokia's research labs.

Todd Jones (published as Stephen) works at Sandia protecting systems from outside attack. He graduated from U of U then went to Wisconsin.

I can't remember the name of the engineer that worked with Kelly when he was at Tandem. He now works at Network Appliance.

Cited Papers that I need to look up

Atomizer (Flanagan '03,'04)

Chen et. al Usenix 2003

Benchmarks used that sounded interesting

Da Capo Blackburn '06 Spec JBB 2000 JVM 98

DBench Mogrify TPC-W SPC Disk Trace

My Random Thoughts

Fake a page fault for an interrupt? Is there something else that would work? Can any interrupt work?