New demonstration CPU miner available

[jgarzik]

Probably because there's a whole JSON parser that's part of the miner...so might as well reuse the code!

Bingo.

[JWU42]

If anybody is interested, this is the assembly code from Intel Compiler which gives me the fastest 4way code on AMD K10. 1.13 GH/s per 1 GHz per physical core.
https://gist.github.com/853566
Compile cpuminer-0.7.1, download the above code, issue:
gcc -c sha256_4way.s
and do make again to link the object file to the executable.  It's about 7% faster than the gcc version.

1.13 GH/s ?!?

[Raulo]

1.13 GH/s ?!?

1.13MH/s, of course. Sorry.

[smgoller]

Just tried to run 0.7.1 on a q6600, ubuntu 10.10 32-bit, and got segfaults when i tried to run cryptopp_asm32. Tried compiling from git, same result. Any thoughts?

[myrkul]

Just tried to run 0.7.1 on a q6600, ubuntu 10.10 32-bit, and got segfaults when i tried to run cryptopp_asm32. Tried compiling from git, same result. Any thoughts?

check your options. I had that happen to me when I had a malformed url in there... mining.bitboin.cz.

[decibel]

Patch for successfull compilation of minerd under FreeBSD:

Expanding on hangover's work; I had to also patch cpu-miner.c and util.c. Complete patch-set is:

Code:

diff -u cpuminer-0.7.1.orig/cpu-miner.c cpuminer-0.7.1/cpu-miner.c
--- cpuminer-0.7.1.orig/cpu-miner.c     2011-02-17 00:54:45.000000000 -0600
+++ cpuminer-0.7.1/cpu-miner.c  2011-03-03 23:38:39.000000000 -0600
@@ -24,7 +24,7 @@
 #include <pthread.h>
 #include <getopt.h>
 #include <jansson.h>
-#include <curl/curl.h>
+#include </usr/local/include/curl/curl.h>
 #include "compat.h"
 #include "miner.h"
 
diff -u cpuminer-0.7.1.orig/miner.h cpuminer-0.7.1/miner.h
--- cpuminer-0.7.1.orig/miner.h 2011-02-17 00:20:34.000000000 -0600
+++ cpuminer-0.7.1/miner.h      2011-03-03 23:28:11.000000000 -0600
@@ -5,7 +5,7 @@
 #include <stdint.h>
 #include <sys/time.h>
 #include <jansson.h>
-#include <curl/curl.h>
+#include </usr/local/include/curl/curl.h>
 
 #ifdef __SSE2__
 #define WANT_SSE2_4WAY 1
@@ -18,7 +18,12 @@
 #if ((__GNUC__ > 4) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3))
 #define WANT_BUILTIN_BSWAP
 #else
-#include <byteswap.h>
+/* #include <byteswap.h> */  // <-- doesn't exist under FreeBSD
+# define bswap_64 __bswap64 
+# define bswap_32 __bswap32 
+# define bswap_16 __bswap16 
+# define __BIG_ENDIAN__ (_BYTE_ORDER == _BIG_ENDIAN)
+#include <sys/endian.h>
 #endif
 
 #if defined(__GNUC__) && (__GNUC__ > 2) && defined(__OPTIMIZE__)
Only in cpuminer-0.7.1: patch
diff -u cpuminer-0.7.1.orig/util.c cpuminer-0.7.1/util.c
--- cpuminer-0.7.1.orig/util.c  2011-03-02 21:06:20.000000000 -0600
+++ cpuminer-0.7.1/util.c       2011-03-03 23:42:56.000000000 -0600
@@ -15,7 +15,7 @@
 #include <stdlib.h>
 #include <string.h>
 #include <jansson.h>
-#include <curl/curl.h>
+#include </usr/local/include/curl/curl.h>
 #include "miner.h"
 
 struct data_buffer {

[adv]

I build debian package 0.7.1 for SPARC architectures (sparc64).

Unfortunately it's more of a test assembly, as on Sun Enterprise 250 (2 * 400MHz) I managed to achieve maximum performance: ~87Khps with "--algo c" and ~81Kps with "--algo cryptopp". Algorithms foк via, 4waySSE2 and cryptopp_asm32 on this architecture is not available. :^(

Some tech info:

Code:

uname -a
Linux debsparc 2.6.32-5-sparc64-smp #1 SMP Wed Jan 12 05:23:55 UTC 2011 sparc64 GNU/Linux

Code:

 cat /proc/cpuinfo
cpu             : TI UltraSparc II  (BlackBird)
fpu             : UltraSparc II integrated FPU
pmu             : ultra12
prom            : OBP 3.22.0 2000/12/20 16:20
type            : sun4u
ncpus probed    : 2
ncpus active    : 2
D$ parity tl1   : 0
I$ parity tl1   : 0
Cpu0ClkTck      : 0000000017d78400
Cpu1ClkTck      : 0000000017d78400
MMU Type        : Spitfire
State:
CPU0:           online
CPU1:           online

Command line:

Code:

btc-cpuminer --url http://192.168.1.1:8334 -r -1 -D -s 5 --threads=2
.....
HashMeter(1): 462143 hashes, 84.75 khash/sec
HashMeter(0): 462143 hashes, 86.89 khash/sec
HashMeter(1): 462143 hashes, 84.75 khash/sec
HashMeter(0): 462143 hashes, 86.80 khash/sec

2 cpu's load at 80-100% by 3 btc-cpuminer threads.

System load: 2.09 2.02 1.74

Is there any hope for improvement in performance for this architecture?


Discussion of issues related to the Debian packages is welcome in this thread: http://bitcointalk.org/index.php?topic=2207

[jgarzik]

Call for hacking:

Anybody want to volunteer to add server failover support to cpuminer?  (m0mchil's miner, too...)

Server failure should not interrupt mining [much], ideally.

[TurdHurdur]

Call for hacking:

Anybody want to volunteer to add server failover support to cpuminer?  (m0mchil's miner, too...)

Server failure should not interrupt mining [much], ideally.

Hah, you're going to make my script do a double-take.

[mtve]

I build debian package 0.7.1 for SPARC architectures (sparc64).

just of curiosity, small performance test here:

sun fire v120 (sun4u), sparcv9 650 MHz - 109.6 Khash/s
sun netra t2000 (sun4v), sparcv9 1.2 GHz - 117.4 Khash/s (strange)

second box has 32 virtual cpus, so supposedly it can do 3.7 Mhash/s total.

yes, cryptopp speed is around 82% of c's.

i've managed to compile only 32bit binary of cpuminer, hope it does not beat performance too much.

Is there any hope for improvement in performance for this architecture?

There is ncp crypto provider in UltraSPARC T1, but it accelerates only RSA (really good).
UltraSPARC T2 with n2cp should help to SHA256.

Would somebody with T2 play with http://frox25.no-ip.org/~mtve/tmp/sun_pkcs11_sha256.c please?   On T1 it's around 58 Khash/s.

Update: from http://cs.anu.edu.au/student/projects/10S2/Reports/Cody%20Christopher.pdf it can be concluded that for 32 bytes (one block of SHA256) and for 64 bytes (full bitcoin block) without special efforts it will be slower than CPU.

[lucky]

0.7.1 fails to build on Mac OS X 10.6.6...

I went with % ./configure ; make

and :

Code:

joel@baldur:~/cpuminer-0.7.1% make
make  all-recursive
Making all in compat
Making all in jansson
gcc -DHAVE_CONFIG_H -I. -I../..     -g -O2 -MT dump.o -MD -MP -MF .deps/dump.Tpo -c -o dump.o dump.c
mv -f .deps/dump.Tpo .deps/dump.Po
gcc -DHAVE_CONFIG_H -I. -I../..     -g -O2 -MT hashtable.o -MD -MP -MF .deps/hashtable.Tpo -c -o hashtable.o hashtable.c
mv -f .deps/hashtable.Tpo .deps/hashtable.Po
gcc -DHAVE_CONFIG_H -I. -I../..     -g -O2 -MT load.o -MD -MP -MF .deps/load.Tpo -c -o load.o load.c
mv -f .deps/load.Tpo .deps/load.Po
gcc -DHAVE_CONFIG_H -I. -I../..     -g -O2 -MT strbuffer.o -MD -MP -MF .deps/strbuffer.Tpo -c -o strbuffer.o strbuffer.c
mv -f .deps/strbuffer.Tpo .deps/strbuffer.Po
gcc -DHAVE_CONFIG_H -I. -I../..     -g -O2 -MT utf.o -MD -MP -MF .deps/utf.Tpo -c -o utf.o utf.c
mv -f .deps/utf.Tpo .deps/utf.Po
gcc -DHAVE_CONFIG_H -I. -I../..     -g -O2 -MT value.o -MD -MP -MF .deps/value.Tpo -c -o value.o value.c
mv -f .deps/value.Tpo .deps/value.Po
rm -f libjansson.a
ar cru libjansson.a dump.o hashtable.o load.o strbuffer.o utf.o value.o 
ranlib libjansson.a
make[3]: Nothing to be done for `all-am'.
gcc -DHAVE_CONFIG_H -I. -pthread -fno-strict-aliasing -I./compat/jansson    -g -O2 -MT cpu-miner.o -MD -MP -MF .deps/cpu-miner.Tpo -c -o cpu-miner.o cpu-miner.c
In file included from cpu-miner.c:29:
miner.h:21:22: error: byteswap.h: No such file or directory
make[2]: *** [cpu-miner.o] Error 1
make[1]: *** [all-recursive] Error 1
make: *** [all] Error 2
joel@baldur:~/cpuminer-0.7.1% 

[jgarzik]

Yep, OSX is known broken at this time.  Waiting for someone to fix it, in a way that doesn't break everyone else's build...

[mrballcb]

Yep, OSX is known broken at this time.  Waiting for someone to fix it, in a way that doesn't break everyone else's build...

There was a guy on IRC last week building your miner who made a quick change per some post he found via google that got byteswap to work.  I'll dig and see if I can find it.

Regards...      Todd

[mrballcb]

Yep, OSX is known broken at this time.  Waiting for someone to fix it, in a way that doesn't break everyone else's build...

There was a guy on IRC last week building your miner who made a quick change per some post he found via google that got byteswap to work.  I'll dig and see if I can find it.

Regards...      Todd

Some asshat came into the channel and trolled the entire freakin channel on Thu and wiped out all my scrollback (gotta remember to increase that on Monday).  But was able to find the post he was talking about.  From http://www.mail-archive.com/barry-devel@lists.sourceforge.net/msg00113.html , there is a link to this commit that defines it inline if OSX is detected:  http://lists.freedesktop.org/archives/xorg-commit/2007-January/010041.html

Regards...     Todd

[fatalfurry]

Is there a way of adding or removing threads on the fly? I like running it with three threads on my 4 core processor but when I'm doing something a little more intense I would want to throttle it down to two or one thread.

[jgarzik]

Version 0.7.2 is released.  Changes:

- Add port of ufasoft's sse2 assembly implementation (Linux only)
  This is a substantial speed improvement on Intel CPUs.

  Big thanks to mmarker/chromicant for this work.

- Move all JSON-RPC I/O to separate thread.  This reduces the
  number of HTTP connections from one-per-thread to one, reducing resource
  usage on upstream bitcoind / pool server.

  This is prepartion for adding long-polling support.

See top of thread for URLs and info.

SHA1: 2673c5f5c2cd85acc332a9c8a1f48415b8dd0fc9  cpuminer-installer-0.7.2.zip
MD5: 24211f26b1d6b0c68fe696662b3082e7  cpuminer-installer-0.7.2.zip

[gbl4ck]

Quick question on estimating my hash rate....  I'm currently running 12 threads and each is running @ ~1400 khash/sec.  Each line looks like:

HashMeter(0): 8288293 hashes, 1455.67 khash/sec
HashMeter(4): 7778834 hashes, 1462.78 khash/sec
HashMeter(11): 8287566 hashes, 1505.17 khash/sec
etc for the rest of the threads (one for each logical processor)

So would my total CPU rate be ~16800 khash/sec?

[jgarzik]

So would my total CPU rate be ~16800 khash/sec?

Yes, that's correct.  HashMeter output is per-thread, not a total for all threads.

Do you really have a 12-core?  If not, I bet you could do better with fewer threads and perhaps another algo.

[gbl4ck]

So would my total CPU rate be ~16800 khash/sec?

Yes, that's correct.  HashMeter output is per-thread, not a total for all threads.

Do you really have a 12-core?  If not, I bet you could do better with fewer threads and perhaps another algo.

Not 12 cores but 6 cores with HT on.  I did just realize tho that running it like this is killing my GPU processing which i'm using poclbm-gui.  I was around 80 Mhash/s but now it drops down to around 30 and will jump to 60 but is very inconsistant. 

[gbl4ck]

Looks like I found a better combo.  I'm still using the 4way algo but I dropped the threads down to 6.  Now all 6 threads are around 2000 khash/s but my GPU is staying steady at 85 Mhash/s.

edit - According to taskmanager i'm only using 50% CPU so maybe I'm not yet in the sweet spot and need to try another algo... i'm running a i7 980x