K6-2/FreeBSD slowness

[alpha]

My K6-2 running FreeBSD 4.7-STABLE seems to be underperforming in my eyes. I don't have any numbers to compare, but this seems a little extreme:

[Mon Mar 10 10:32:43 2003] iteration: 20000/3102083 (0.64%) k = 21181 n = 3102068
[Mon Mar 10 12:00:05 2003] iteration: 30000/3102083 (0.97%) k = 21181 n = 3102068
[Mon Mar 10 13:27:28 2003] iteration: 40000/3102083 (1.29%) k = 21181 n = 3102068
[Mon Mar 10 14:56:26 2003] iteration: 50000/3102083 (1.61%) k = 21181 n = 3102068

Now I've read most of the stuff posted on this forum, so I realise the k and/or(?) n values will have an affect, but do these numbers look reasonable? Does anyone have similar hardware?

I guess the K6-2 lacks SSE and SSE2, could this be a (partial) reason?

Oh and before someone suggests it - that box has been finely tuned and is running a stripped down kernel. At the time of the above, X was not running nor were any time-consuming daemons/background apps.

Edit: the client is v1.0.2.

[mklasson]

I'm sorry to say this, but those numbers look about right to me. I don't know what frequency your k6-2 runs at, but I figure it's around 300.

The K6-2 really is horrendously slow. Sorry. And yes, they do lack SSE and SSE-2. They do have 3dNow!, but that doesn't help much in this scenario, methinks. The prefetch instruction there might give it a slight boost, but not enough to exit the Pretty Worthless region.

Good thing a cpu that's 20 times faster only costs about $50 nowadays.

[alpha]

Originally posted by mklasson
I'm sorry to say this, but those numbers look about right to me. I don't know what frequency your k6-2 runs at, but I figure it's around 300.

I knew I forgot something Its a 400MHz CPU in actual fact.

The K6-2 really is horrendously slow. Sorry. And yes, they do lack SSE and SSE-2. They do have 3dNow!, but that doesn't help much in this scenario, methinks. The prefetch instruction there might give it a slight boost, but not enough to exit the Pretty Worthless region.

Fair enough. Its a matter of weighing up which client for which project is most efficient for this CPU. It actually fares quite well at distributed.net's OGR projects. ~2.5 Mnodes/sec while my XP 1700 does around 11.5 Mnodes/sec.

Good thing a cpu that's 20 times faster only costs about $50 nowadays.

Indeed, but I don't take DC seriously enough to upgrade machines simply for that purpose. The box in question runs FreeBSD smoother than I could ask for, and I am more than happy with its performance.

[Troodon]

- The K6 family has the lowest latency FPU in the world but it isn't pipelined. This doesn't help in a world optimized for pipelined FPUs.
- Since it's a low latency non-pipelined FPU it also needs to be well feeded with instructions, which makes very important the caches and the memory subsystem.
The K6 and the K6-2 have a 64 kB L1 cache, but they don't have integrated L2 cache. The cache on the motherboard doesn't help much because is high latency, shares bandwidth with the rest of data that uses the FSB, and works at the FSB speed (100 MHz for the K6-2).
If you look at the K6-III or the "+" family (which have integrated L2 cache), at the same speed, they perform better.
- Plus, the PRP testing depends mostly on the caches and the memory subsystem.
- The K6 family "shines" only at integer operations.

Conclusion/advices:
- The K6 family will perform very slow at floating point operations, especially on those involving double (80 bit) precision.
- The K6 family will perform very slow at calculations that rely on the memory subsystem.
- Don't use your K6 for PRP testing.
- Distributed computing projects where the K6 family performs well: OGR and Euler2000.
- The best way of overclocking a K6 family member is by rising the FSB.

[shauge]

I run WinXP on a K6-3 450MHz, the last proth test was completed in 17 days and 8 hours. Compared to my Athlon 650MHz, that completes a proth test in around 3 days and 12 hours, it seem as if the K6-3 comes short in this project. I also have the experience that it does comparativly much better in other projects.

Edited:
I agree with Troodon's conclusions.

[alpha]

Originally posted by Troodon
- Don't use your K6 for PRP testing.
- Distributed computing projects where the K6 family performs well: OGR and Euler2000.
- The best way of overclocking a K6 family member is by rising the FSB.

Wow, firstly thanks for the in-depth post, I didn't really know much about the K6-2 until now.

I've taken the machine off SoB and put it back on OGR. I've been interested in OGR for a while now, I was just 'dabbling' with SoB.

: I'm an overclocking newbie to say the least. The only successful overclock I made was a p100 @ 120MHz, but it was unstable due to lack of cooling. Whenever I've tried to overclock the K6-2 it fails to POST. I'm not entirely sure, but I can only remember fiddling with the multiplier - is this why? Should I be fiddling with the FSB too? Or am I getting my jargon fuddled

[alpha]

Originally posted by shauge
I also have the experience that it does comparativly much better in other projects.

Maybe you could put that puppy on OGR, we could use a little help

[MAD-ness]

To get a rough estimate of SoB performance across various architectures, you could look at the GIMPS (Prime95) benchmarking page.

While performance will not be identical (or perhaps even all that close), RELATIVE performance, atleast in regards to architecture and instruction set optimization, should be pretty similar.

http://www.mersenne.org/bench.htm

Type Speed
(MHz) Memory
Speed L2
Cache
Size L2
Cache
Speed 6.52M
to
7.76M
(384K) 7.76M
to
9.04M
(448K) 9.04M
to
10.33M
(512K)

AMD K6-2 400 100 1024 Bus 0.529 0.640 0.708

Celeron 400 66 128 Full 0.235 0.282 0.315 

P-II 400 100 512 Half 0.207 0.247 0.276

Formatting sucks, but it is better than nothing.

Smaller iteration times are better (faster).

At the given (smallest included) exponent size and FFT size:

AMD K6-2 400                 0.529
Celeron 400                    0.235
P-II 400                           0.207

While the FFT algorithms and code hit the system bus hard, love low latency and high speed memory and are also cache-dependent, the type of work being done is very FPU intensive.


At 20.4M to 25.35M sized exponents
(1280K FFT size):

AMD K6-2 400             2.424
Celeron 400                0.917
P-II 400                      0.769

At this point, the FFT is quite obviously too big for the entire FFT to fit into the cache of any mainstream CPUs.

Here we see the Pentium 2 core CPUs just destroy the K6-2 cored CPU.

[alpha]

Thanks for the info MAD-ness. I now know how truly useless my K6-2 is. I'm just going to leave it quietly crunching OGR and never mention it again :P

[MAD-ness]

To the best of my knowledge, there isn't much reason to run K6 computers on anything other than OGR. It is the only project that actually does as well or better on the K6 architecture as on other architectures.

Also, it isn't USELESS, but like many of the Via chips (such as the C3) it is so crippled in terms of FPU power that it performs pretty badly on most DC projects.

[shauge]

For whom it may interest: I tried my two computers on the OGR project. The performance ratio between the AMD K7 Athlon 650MHz and the AMD K6-III 450MHz dropped from about 5 in this project, to about 1.7. That is just slightly more than the ratio between the clock frequences, about 1.45.

[Troodon]

Alpha, maybe your K6-2, like the mine one, needs 2.3 or 2.4 V to work at 450 MHz. Another thing you can try is to lower the multiplier and raise the FSB.

[alpha]

Originally posted by Troodon
Alpha, maybe your K6-2, like the mine one, needs 2.3 or 2.4 V to work at 450 MHz. Another thing you can try is to lower the multiplier and raise the FSB.

I'll have another fiddle as soon as I get the chance, only thing is I have no idea how hot the thing is getting I'll post any success/failures in this thread.

[alpha]

Originally posted by Troodon
Alpha, maybe your K6-2, like the mine one, needs 2.3 or 2.4 V to work at 450 MHz. Another thing you can try is to lower the multiplier and raise the FSB.

OK I had some free time today (gf hijacked main machine to play AOE ) so I tried some overclocking.

Firstly, the motherboard manual only lists 2.1, 2.2, 2.5, 2.7, 2.8, 2.9, 3.2, 3.3 and 3.5 for the voltage. I believe the K6-2 400 is supposed to run at 2.2V, so I left that alone to begin with. Before I fiddled, the multiplier was 4x, and the FSB was 100MHz "SDRAM 100M" (max).

I tried the following combinations:

83MHz 5.0x = 415MHz
83MHz 5.5x = 457MHz
95MHz 4.5x = 428MHz
95MHz 5.0x = 475MHz
95MHz 5.5x = 523MHz

There were two 95MHz available, one was "SDRAM 95M", and the other was "66M". I just used the 95M one.

All of the above were unstable, or caused the machine not to POST. I thought all was well at 415MHz, but after about 10min of dnetc action, it kernel panicked (others panicked on boot, or seconds after booting). I assume the instablity was due to overheating, but I was a little surprised since that box is always ice cool and has a good HSF.

Anyway, after this I toyed with the voltages a little - by bumping it up to 2.5V for 457MHZ and a couple of others, but everything kept bailing out as before.

So - is this due to overheating or am I doing something wrong? Any tips? I feel quite unfulfilled after fiddling with jumpers for an unmentionable period of time

[smh]

So - is this due to overheating or am I doing something wrong?

Most probably it's caused by overheating.

I remember, years ago, that we had to replace dozens of cpu fans and heatsinks because the ones that were in there weren't sufficient and caused intermittend problems on > 80% of the pc's. Replacing the fans resolved almost all problems

[Troodon]

Alpha, is There termal compound between the CPU and the heatsink?
Have you tried to set 4.5x100 @ 2.5 V?
I don't recommend you to set the FSB @ 83 MHz, because then the PCI bus runs at 83/2 MHz. The nominal PCI speed is 33 MHz and at 41,5 MHz your hard drive or any other PCI device can fail.

[alpha]

Originally posted by Troodon
Alpha, is There termal compound between the CPU and the heatsink?

If you mean some kind of thermal paste - no.

Have you tried to set 4.5x100 @ 2.5 V?

No - I don't think so. I'll give it a try, but I figured if 415MHz was causing overheating, anything higher would too.

I don't recommend you to set the FSB @ 83 MHz, because then the PCI bus runs at 83/2 MHz. The nominal PCI speed is 33 MHz and at 41,5 MHz your hard drive or any other PCI device can fail.

OK. Thanks for the hints - I'll let you know what happens with 4.5x100@2.5V (tomorrow).

[Troodon]

Before anything else, apply some thermal grease. If your heatsink has a thermal pad, remove it and leave only the grease.
The difference between 400 and 415 MHz is too small to cause overheating. The problem was with the 41 MHz PCI bus.
If your motherboard doesn't support more than 100 MHz FSB, use only that speed. 83x5.5 will give you less performance than 100x4.5.

[alpha]

Originally posted by Troodon
Before anything else, apply some thermal grease. If your heatsink has a thermal pad, remove it and leave only the grease.
The difference between 400 and 415 MHz is too small to cause overheating. The problem was with the 41 MHz PCI bus.
If your motherboard doesn't support more than 100 MHz FSB, use only that speed. 83x5.5 will give you less performance than 100x4.5.

OK, I tried 4.5x100@2.5V and the machine was stable for a few minutes but then hung a couple of times. I guess that was due to overheating which didn't really surprise me. I'm not too bothered about this, just wondered if I could squeeze anything extra from it without additional cost. Thanks for the help and advice.

Edit: fixed drunken grammar.