Re: [Gems-users] Physical Memory Parameters


Date: Fri, 22 Jul 2005 15:41:16 +0100 (BST)
From: Brinda Ganesh <brinda_ganesh@xxxxxxxxx>
Subject: Re: [Gems-users] Physical Memory Parameters
Hi

Im setting g_NUM_MEMORIES 1 in the
rubyconfig.defaults.
and the Network FILE_SPECIFIED as well.

Im using a modified version of the MOESI_SMP_directory
protocol which redirects all memory requests to my
dram simulator.

I will take a look at the CMP protocol.

Thanks
Brinda
--- Bradford Beckmann <beckmann@xxxxxxxxxxx> wrote:

> 
> Brinda,
> 
> I'm confused.  First, you don't seem to be setting
> the "-m
> <number_of_memories>" command line parameter so
> config/RubyConfig.C should
> be setting g_NUM_MEMORIES to be equal to the
> g_NUM_PROCESSORS, which is
> the default.  However, you say the file you are
> using is named:
> NUCA_Procs-2_ProcsPerChip-1_L2Banks-2_Memories-1.txt
> , but I would have
> expected network/simple/Topology.C to load the file
> named:
> NUCA_Procs-2_ProcsPerChip-1_L2Banks-2_Memories-2.txt
> because
> g_NUM_MEMORIES == 2.  Are you sure that is the
> network file being loaded?
> 
> Second, are you trying to use your FILE_SPECIFIED
> network using the
> MOSI_SMP_bcast protocol?  The broadcast snooping
> protocol won't work
> because the FILE_SPECIFIED networks don't provide
> the ordering required by
> the protocol.
> 
> If you want to have only one memory in the system, I
> suggest you use one
> of the CMP directory protocols (easy solution), or
> make the necessary
> edits to the HIERARCHICAL_SWITCH network creation
> code in Topology.C and
> possibly some slightly changes to the MOSI_SMP_bcast
> .sm files (harder
> solution) to support hierarchal networks where
> processors != memories.
> 
> Brad
> 
> 
>  > The network setup that I have currently looks
> like
> > this
> > is named
> >
> NUCA_Procs-2_ProcsPerChip-1_L2Banks-2_Memories-1.txt
> > The file contents are as below
> >
> > processors:2
> > procs_per_chip:1
> > L2banks:2
> > memories:1
> > bw_unit:10000
> >
> > ext_node:L1Cache:0 int_node:0 link_latency:1
> > bw_multiplier:64
> > ext_node:Directory:0 int_node:0 link_latency:40
> > bw_multiplier:10
> > int_node:0 int_node:1 link_latency:40
> bw_multiplier:16
> >
> > ext_node:L1Cache:1 int_node:1 link_latency:1
> > bw_multiplier:64
> > ext_node:Directory:1 int_node:1 link_latency:40
> > bw_multiplier:1
> >
> > The last Directory's link presence causes the
> first
> > error I mentioned in my earlier email while its
> > absence causes an error later on.
> >
> > This file is only for the 2 processor case - but I
> > definitely would like to extend it to the 4/8
> > processor case as well.
> >
> > The arguments that I use for the tester are
> >
> > tester.exec -p 2 -l 500 -v med -s n
> >
> > Thanks
> > Brinda
> >
> >
> > --- Bradford Beckmann <beckmann@xxxxxxxxxxx>
> wrote:
> >
> > >
> > > Brinda,
> > >
> > > Mike is right.  The HIERARCHICAL_SWITCH network
> > > assumes that the number of
> > > memories equals the number of chips.  This is
> > > because the hierarchical
> > > switch builds an ordered network among multiple
> > > chips assuming each chip
> > > has a separate memory controller.  Actually,
> this
> > > probably is something we
> > > can easily fix in our next release, but
> currently
> > > GEMS is limited that
> > > all auto-generated networks must have a number
> of
> > > memories >= the number
> > > of chip.
> > >
> > > As far as your problem with the FILE_SPECIFIED
> > > network, are you trying to
> > > create a network with more directories but the
> > > g_NUM_MEMORIES equals one?
> > > Can you send us the file you are using and the
> > > command line arguments you
> > > are passing the tester?
> > >
> > > Brad
> > >
> > >
> > >
> > > On Thu, 21 Jul 2005, Brinda Ganesh wrote:
> > >
> > > > I created a network file similar to the ones
> in
> > > the
> > > > network/simple/Netowrk_Files. I set the number
> of
> > > > memories to 1 in this file. I am using  2
> > > processors
> > > > and a modified form of the MOESI_SMP_directory
> > > > protocol.
> > > >
> > > > The netowrk type is specified as
> FILE_SPECIFIED.
> > > >
> > > > The network initialization crashes if I
> specify a
> > > link
> > > > for more than 1 Directory.
> > > > failed assertion 'nodes[i] >= 0 && nodes[i] <
> > > m_nodes'
> > > > at fn void Topology::makeFileSpecified() in
> > > > network/simple/Topology.C:511
> > > >
> > > > If I specify only one directory I get the
> crash
> > > that
> > > > occurred earlier with the Hierarchial switch
> > > network
> > > > type.
> > > > failed assertion 'id < m_nodes' at fn void
> > > > SimpleNetwork::checkNetworkAllocation(NodeID,
> > > bool,
> > > > int) in network/simple/SimpleNetwork.C:199
> > > >
> > > > This seems to be related to a call from
> > > > getToNetworkQueue being called during chip
> > > > initialization.
> > > >
> > > > Im a little confused about whether its the way
> my
> > > > network file is being specified which is
> causing
> > > the
> > > > error or if there is some other parameter I
> need
> > > to
> > > > change as well.
> > > >
> > > > Thanks
> > > > Brinda
> > > >
> > > > --- Brinda Ganesh <brinda_ganesh@xxxxxxxxx>
> wrote:
> > > >
> > > > > Im using the HIERARCHICAL_SWITCH network.
> > > > > Thanks
> > > > > Brinda
> > > > >
> > > > > --- Mike Marty <mikem@xxxxxxxxxxx> wrote:
> > > > >
> > > > > > > How does ruby pass back the installed
> memory
> > > > > size
> > > > > > > information back to simics? Im assuming
> that
> > > > > this
> > > > > > > information is conveyed somehow so that
> the
> > > os
> > > > > can
> > > > > > > swap things in and out of disk if the
> > > virtual
> > > > > > memory
> > > > > > > size gets larger than the installed
> physical
> > > > > > memory
> > > > > > > size.
> > > > > > >
> > > > > >
> > > > > > Ruby doesn't pass this information to
> Simics.
> 
=== message truncated ===



		
___________________________________________________________ 
How much free photo storage do you get? Store your holiday 
snaps for FREE with Yahoo! Photos http://uk.photos.yahoo.com
[← Prev in Thread] Current Thread [Next in Thread→]