jackin@vehka.uta.fi (Markku M{enp{{) (11/01/90)
I am wondering about how faster ram affects perfomance.
I understand that non-cache boards probably benefit from
faster ram but is it the same with cache MBs too ?
Say for example if I have a 33Mhz/64K cache board, do
I get a speed increase if I use 70ns ram instead of 80ns.
It seems that currently ram write is 1 ws and read 0 ws.
Markku
--
Markku M{enp{{
email : jackin@vehka.uta.fi
phone : 358 31 561 575 (in Finland (931) 561 575)silver@xrtll.uucp (Hi Ho Silver) (11/05/90)
In article <1718@kielo.uta.fi> jackin@vehka.uta.fi (Markku M{enp{{) writes:
$I am wondering about how faster ram affects perfomance.
$I understand that non-cache boards probably benefit from
$faster ram but is it the same with cache MBs too ?
Yes, but nowhere near as much. The idea of a cache is that the vast
majority of memory references occur in a fairly small area of memory
during any given time period. For example, in a matrix multiplication
routine, you'd have some loop code that gets executed repeatedly, and
your data fetches into the matrices would also be done repeatedly. By
loading these frequently-accessed areas into a fast cache, your computer
doesn't have to wait for main memory as often.
If 90% of your memory accesses are serviced out of the cache, and your
main memory runs at 2 wait states (and assuming cache misses take no
additional clock cycles of their own), your average number of wait states
would be
.9 * 0 + .1 * 2 = .2
(the cache) (main memory)
If you reduce your main memory wait states to 1 by replacing the
chips with faster memory, you then end up with an average of .1 wait
states per memory access.
On my old 286 system, running at 0 ws was about 10% faster than
running at 1 ws; if this holds for 386 systems as a guideline, then
in the above example, you'd only speed your machine up by 1%. Admittedly,
the numbers are entirely made-up, but they should show you that you are
not likely to gain much by speeding up main memory on a cached motherboard.
Keep in mind that although a 90% hit rate may sound pretty high, you
actually don't need a very large cache to achieve that kind of rate. I
don't have the figures available, but that rate should be achievable with
a 64K cache, and perhaps even with a 32K one.
--
HI ROGER |Nikebo says "Nikebo knows how to post. Just do it."| silver@xrtll
_________|-----------------------|_______________|------------|_____________
yunexus!xrtll!silver (L, not 1) | Hi Ho Silver | costing the net thousands
Silver: Ever Searching for SNTF |i need a grilf | upon thousands of dollarsjackin@vehka.uta.fi (Markku M{enp{{) (11/05/90)
In article <1990Nov4.221653.4823@xrtll.uucp> silver@xrtll.UUCP (Hi Ho Silver) writes: >In article <1718@kielo.uta.fi> jackin@vehka.uta.fi (Markku M{enp{{) writes: >$I am wondering about how faster ram affects perfomance. >$I understand that non-cache boards probably benefit from >$faster ram but is it the same with cache MBs too ? > > Yes, but nowhere near as much. > If 90% of your memory accesses are serviced out of the cache, and your >main memory runs at 2 wait states (and assuming cache misses take no >additional clock cycles of their own), your average number of wait states >would be > > .9 * 0 + .1 * 2 = .2 > (the cache) (main memory) > > If you reduce your main memory wait states to 1 by replacing the >chips with faster memory, you then end up with an average of .1 wait >states per memory access. This is true for write-back cache method. It doesn't update main memory if a write operation remains in cached areas. This method is used in for example Everex machines. However the most common method is write-through method which *allways* updates main memory when a write operation occures. So if our read/write-operations go like this : 90% hits cache 10% misses cache .45 * 0 + .45 * 2 + .1 * 2 = 1.1 (cache/read) (cache/write) (main memory) Now if we could reduce our main memory waitstates by 1 we could get .55 waitstates which could give us a noticable speed increase over 1.1 ws. The third method, posted write-through cache, is also used but I don't know if it is very common. Markku -- Markku M{enp{{ email : jackin@vehka.uta.fi phone : 358 31 561 575 (in Finland (931) 561 575)