linux-origin
[Top] [All Lists]

Cache tuning results

To: Kanoj Sarcar <kanoj@xxxxxxxxxxxx>
Subject: Cache tuning results
From: Ralf Baechle <ralf@xxxxxxxxxxx>
Date: Sun, 23 Jul 2000 02:08:34 +0200
Cc: linux-origin@xxxxxxxxxxx
Sender: owner-linux-origin@xxxxxxxxxxx
This are lmbench results from recent Linux/MIPS kernels.  The machines

 - vitima is Origin 200, 2 x R10000, 1mb L2 running IRIX 6.5.4
 - ralf is Origin 200, 2 x R10000, 2mb L2 running Linux 2.4.0-test5-pre3
   with cache and TLB optimizations.
 - ice is an R4400 200MHz, 1mb L2 Indy running 2.4.0-test5-pre1 stock CVS
   kernel.
 - Ulf results where obtained by Ulf on penguin4 with 2.3.99-pre8.

Note the dramatical improvments of the ``exec proc'' benchmark due to the
cache hacks.  The latency benchmarks are entirely being won by Linux, the
local bandwith benchmarks are still all won by IRIX - mosty because our
memcpy and ip checksum code are still just 32-bit code.  The ridiculously
bad mmap latency numbers for `ralf' are due to running diskless.  The numbers
for penguin4 show how Linux actually performs.

Processor, Processes - times in microseconds - smaller is better
----------------------------------------------------------------
Host                 OS  Mhz null null      open selct sig  sig  fork exec sh  
                             call  I/O stat clos       inst hndl proc proc proc
--------- ------------- ---- ---- ---- ---- ---- ----- ---- ---- ---- ---- ----
vitima       IRIX64 6.5  180  2.8  11.  103  132 0.27K  8.1   29 1.4K   8K  14K
ralf      Linux 2.4.0-t  180  1.2  1.8   20   26 0.13K  3.0   11 0.5K   1K  36K
ice       Linux 2.4.0-t  198  0.9  2.0   25   28 0.15K  6.6   16 4.9K  19K  64K
ulf       Linux 2.3.99-  180  1.3  1.8   13   15 0.13K  3.0   15 1.2K  13K  43K

Context switching - times in microseconds - smaller is better
-------------------------------------------------------------
Host                 OS 2p/0K 2p/16K 2p/64K 8p/16K 8p/64K 16p/16K 16p/64K
                        ctxsw  ctxsw  ctxsw ctxsw  ctxsw   ctxsw   ctxsw
--------- ------------- ----- ------ ------ ------ ------ ------- -------
vitima       IRIX64 6.5   18     19     44    31     50      61     137
ralf      Linux 2.4.0-t    2      6     42    15     48      18      95
ice       Linux 2.4.0-t    4     78    529    91    402     115     600
ulf       Linux 2.3.99-    1      6    216    46    582     123     966

*Local* Communication latencies in microseconds - smaller is better
-------------------------------------------------------------------
Host                 OS 2p/0K  Pipe AF     UDP  RPC/   TCP  RPC/ TCP
                        ctxsw       UNIX         UDP         TCP conn
--------- ------------- ----- ----- ---- ----- ----- ----- ----- ----
vitima       IRIX64 6.5    18    76  136   187         194        382
ralf      Linux 2.4.0-t     3    14   31    54          78        251
ice       Linux 2.4.0-t     4    24   66
ulf       Linux 2.3.99-     1    16   38

File & VM system latencies in microseconds - smaller is better
--------------------------------------------------------------
Host                 OS   0K File      10K File      Mmap    Prot    Page
                        Create Delete Create Delete  Latency Fault   Fault
--------- ------------- ------ ------ ------ ------  ------- -----   -----
vitima       IRIX64 6.5                                10305     7   16.2K
ralf      Linux 2.4.0-t                               531477     2    0.0K
ice       Linux 2.4.0-t     53      5    103     11     4034     4    0.0K
ulf       Linux 2.3.99-     24      4     62     14     7443     2    0.0K
[...]

  Ralf

<Prev in Thread] Current Thread [Next in Thread>
  • Cache tuning results, Ralf Baechle <=