bernhold@red (David E. Bernholdt) (03/31/91)
In article <1991Mar30.142903.5225@ariel.unm.edu> prentice@triton.unm.edu (John Prentice) writes: >One other question. What is the size of the memory banks on the >YMP and hence what strides should you avoid? I asked this question a while ago. I _think_ that the (two) YMPs to which I have access have 256 banks. In general, there are a wide range of possible bank configurations for a given type of Cray. Unicos has a command and a system call called "target" which report type of Cray, memory size, number of banks, clock period, etc. for the host in question. The system call has to be made via a C routine, but in Unicos it is fairly easy to call C from Fortran. Once you've got the number of banks, the most general thing to do is to work with vectors who's length is prime relative to the number of banks. One caveat on the "target" command/call: the reponse returned can be configured by the system manager, though I'm not sure I see the value in "lying" about the machine's configuration. -- David Bernholdt bernhold@qtp.ufl.edu Quantum Theory Project bernhold@ufpine.bitnet University of Florida Gainesville, FL 32611 904/392 6365
djh@xipe.osc.edu (David Heisterberg) (03/31/91)
In article <27758@uflorida.cis.ufl.EDU> bernhold@red (David E. Bernholdt) writes >I asked this question a while ago. I _think_ that the (two) YMPs to >which I have access have 256 banks. >... >Once you've got the number of banks, the most general thing to do is >to work with vectors who's length is prime relative to the number of >banks. For single stream performance it is the number of sub-sections rather than banks that is important. A memory reference makes a particular bank unavailable to all CPUs, and the containing sub-section unavailable to further references from the same CPU, for 5 CP. I lack a definative statement, but I think all YMPs have 32 sub-sections. The 2 series has 64 banks; the 4, 128; and the 8, 256. -- David J. Heisterberg djh@osc.edu And you all know The Ohio Supercomputer Center djh@ohstpy.bitnet security Is mortals' Columbus, Ohio 43212 ohstpy::djh chiefest enemy.
ftower@ncar.ucar.EDU (Francis Tower) (04/02/91)
John,
I was out to lunch on my memory stride comments. David Heisterberg
(a nice german name) was correct. It is subsection conflicts which
have precedence over the bank conflicts for a single CPU. The memory
stride and relative performances for the subsections are:
Mod(STRIDE, 32) rel. Performance
0 1/5
1 - 7 1
8 4/5
9 - 15 1
16 2/5
17- 23 1
24 4/5
25-31 1
There are 3 types of memory conflicts:
Section -- two or more ports in the same CPU want to access the
same memory section.
Subsection -- a port in the same CPU wants to access the same
subsection referenced by another port. Wait time is
1 to 4 clocks.
Bank -- any port from any CPU wants to access a bank (at the
same time or subsequently). Wait time is 1 to 5 clocks.
Cray breaks this into two cases 'Simultaneous Bank' and
'Bank Busy'.
Today, It's Racquet Ball!
Francis G Tower
Software QA
NCAR/CGD/ICS
<< Middle-aged Mutant Ninja Modelers >>
"Don't be misled by truth. Science is fact!"