🚀 go-pugleaf

In article <1e824rINNlpu@iraul1.ira.uka.de> wolpers@i11s10.ira.uka.de (Andreas Wolpers) writes:
>Hello everybody,
>
>I'm having a minor problem on which I would welcome any comments:
>Since I'll have some money to spend real soon now, I've been pondering
>whether we should switch from Sun to HP. The usual benchmark results
>suggest that a change might result in faster execution of our pet
>program, a large "theorem prover" written in Lisp.
>
>From the SpecInt92 results, one should expect a performance increase
>of about a factor of 2 when switching from a SS2 to either s SS10-30
>or HP720. Unfortunately, on both machines turned out to but just
>30% faster than a SS2 when running our system (for which ps NEVER
>shows a resident set under 3-4 MB, not even on an 8MB machine).
>
>Any explanations at hand? Did we encounter a bottleneck between
>CPU and memory, or what?

Could be.  If you don't have a generational GC, your locality is
going to be the pits.  If you allocate a lot of data between
garbage collections, you'll typically incur a cache miss and
a writeback for every block of memory you allocate.  That's
because you can't reuse memory until you know it's garbage,
so you're always allocating something you haven't used for
a long time, i.e., at least since the last garbage collection.

What you want is a generational garbage collector and a cache
large enough to hold the youngest generation.  This lets you
allocate less-than-a-cache-full of data between garbage
collections, reclaim most of the space, and reuse it at
the next gc cycle.

The youngest generation should generally be >100KB for basic
GC efficiency reasons (space-time tradeoffs), so you can't
really expect to stay in a first-level cache for your heap
allocations.  You could stay in a megabyte-range second
or third level cache.

> What performace should I expect from
>a SS10-41 or SS10-52 (which have a larger cache, but still not 
>large enough to hold the Lisp's resident set. And if I'm not
>mistaken, the large cache results in a longer time spent for
>non-cache memory accesses (6 cycles instead of 3)). What 
>performace should I expect from a HP735?

If cache misses on allocation are your problem, you're limited
by the rate of allocation and the cache miss service time,
plus something for write backs.  (You'll typically incur a
write-back of a dirty block for each block of heap data you
allocate, since the cache will be mostly full of relatively
recently allocated---hence written--garbage.  This can overload
your write buffers in a hurry for some programs.)

You also need to add something extra if it's a direct-mapped
cache---gc'd systems are especially sensitive to DM cache
conflicts.  (Actually, it's kinda weird--DM works BETTER
if the youngest generation almost fits in the cache, but
not quite.)

So if you know the rate of allocation in your application,
you should be able to figure a ballpark cache miss cost
without much trouble.

(For more on this, see Wilson, Lam, and Moher, "Caching Considerations
for Generational Garbage Collection," ACM Lisp & Functional Programming
'92.)

>I can buy enough memory so that disk speed is no criterion.
>Should I wait for machines with 8MB of cache? Should I shoot myself?
>Should I give YOU the money? :-)

You might want to look for a better Lisp system instead, with
a generational GC.  Higher cache-to-memory bandwidth could be a 
big win, or even just a bigger block size.  (It depends on whether 
you're running up against bandwidth problems or latency problems, 
but 32- or 64-byte blocks are probably considerably better than 16.
Prefetching would probably work even better than large block sizes,
up to the point you're bandwidth-limited.)

If you DO have a generational GC, it's important to have high
bandwidth to any caches that won't hold the youngest generation,
and it's good to have a great big secondary cache that will.

And yes, you definitely should send me the money, if there's any 
left over after buying a machine with a MB or so of cache.

    -- Paul

-- 
| Paul R. Wilson                                    wilson@cs.utexas.edu      |
| U. of Texas Computer Sciences Dept.               voice: (512) 471-9555     |
| Taylor Hall 2.124, Austin, TX 78712-1188          fax:   (512) 471-8885     |
| "Inertia makes the world go 'round."                                        |