Jim, I see you got a Ryzen 3700x. I'm very curious to see how that pans out for you.

I had high hopes for it. It started out very well, around 1200 points for the 24-hour work units. Then after a few days it dropped down to around 200 points, so I stopped.

Ugh. I'm looking at this for my next system. I get more than that now using an FX8370

---

Jim, I see you got a Ryzen 3700x. I'm very curious to see how that pans out for you.

I had high hopes for it. It started out very well, around 1200 points for the 24-hour work units. Then after a few days it dropped down to around 200 points, so I stopped.

Ugh. I'm looking at this for my next system. I get more than that now using an FX8370

Sometimes the shorter (12 or 18 hour) work units do better. I think the scoring system is off track. The Ryzen 3700x is a nice chip, and there are plenty of other uses for it if it doesn't work out for Rosetta.

---

Jim, I see you got a Ryzen 3700x. I'm very curious to see how that pans out for you.

I had high hopes for it. It started out very well, around 1200 points for the 24-hour work units. Then after a few days it dropped down to around 200 points, so I stopped.

Ugh. I'm looking at this for my next system. I get more than that now using an FX8370

Sometimes the shorter (12 or 18 hour) work units do better. I think the scoring system is off track. The Ryzen 3700x is a nice chip, and there are plenty of other uses for it if it doesn't work out for Rosetta.

The credit system here makes 0 sense to me. It seems like if a workunit of the same type is completed by quite a few people, the credit given is drasttttttically reduced, regardless of the machine you're on, which makes it hard to benchmark and very frustrating especially since it makes one wonder if their CPU cycles are actually contributing to anything or if you're just in limbo.
If given the option, I'd love Rosetta to switch to something like CPDN, where it may take a few days, but you have the trickle up system, meaning you still get credit gradually. Either that or just have workunits be workunits, without a target time, which hardly seems to do anything but change the amount of tasks you download and, honestly, the shorter time probably means higher probability that you get a task that hasn't been completed by a lot of other machines, thus granting you more credit.

I'm overcomplicating things here obviously but this is one of the more frustrating projects to get a handle on, especially with consistent credit and figuring out if you are actually doing anything. Since an i7-3770 can beat my Ryzen core for core, credit for credit. That hardly seems right.

The credit system here makes 0 sense to me. It seems like if a workunit of the same type is completed by quite a few people, the credit given is drasttttttically reduced, regardless of the machine you're on, which makes it hard to benchmark and very frustrating especially since it makes one wonder if their CPU cycles are actually contributing to anything or if you're just in limbo.

Me too. But I originally got about 200 points per WU on the first set of work units on my i7-9700 (Ubuntu 18.04) when running the 24-hour work.
I just tried the 12-hour work units, and got 977.04 points. I like that better.
https://boinc.bakerlab.org/rosetta/results.php?hostid=3735813

It is not the credit per se, but I have no idea what is actually going on. I would think a big (BIG) project like Rosetta could do better.

---

The credit system here makes 0 sense to me. It seems like if a workunit of the same type is completed by quite a few people, the credit given is drasttttttically reduced, regardless of the machine you're on, which makes it hard to benchmark and very frustrating especially since it makes one wonder if their CPU cycles are actually contributing to anything or if you're just in limbo.

Me too. But I originally got about 200 points per WU on the first set of work units on my i7-9700 (Ubuntu 18.04) when running the 24-hour work.
I just tried the 12-hour work units, and got 977.04 points. I like that better.
https://boinc.bakerlab.org/rosetta/results.php?hostid=3735813

It is not the credit per se, but I have no idea what is actually going on. I would think a big (BIG) project like Rosetta could do better.

Combine with the errors I see happening (mostly because I ran out of memory, I think) I have set this project to 0% while I crunch away at Asteroids. Losing hours of CPU work for not even a little credit is disheartening. However AVX tasks on asteroids, even on supported processors rarely gets sent, so that's yet another problem, and Rosetta isn't even remotely close to optimized for that (I think maybe SSE 2 or 3, maybe) not even that? It's like they just don't care about the enormous gains in efficiency and our electric bill vs. work done that they could gain.


That being said, I haven't given up entirely. I set project switching to 1200 in preferences and checkpointing to 120. I'm not sure what a good number is. I also set boinc to use 20^ of virtual memory - before it was set to an absurd amount - 75% - so it's no wonder tasks were stuck thrashing and eventually erroring out. Not sure why Boinc feels that is acceptable.
I think Asteroids is soon to run out of work for a little while so I'll have more time to experiment what works in Rosetta and what does not. It's a far cry from World community grid, where I have been exclusively crunching for nearly 2 years, and only now ventured out to try other projects. It is very much set it and forget it.

I recently built a 3600X and it was working out at 14k per day with 12 threads, or 12k with 6 threads.

I built it for family so it's not running that much now that I've handed it over, but it's here:

https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=3728342

One thing that Is really like to see the CPUs or Windows (not sure which manages clock speed these days) do better is manage the clock speed when running low priority tasks like BOINC projects. Ideally they would run at the base clock speed and then the CPU would still be able to turbo for other higher priority tasks. It's fine if you've got great cooling, but my last build was a small desktop with the stock fan so it got quite loud under load. Anyone know if there's a way to make that happen?

---

I recently built a 3600X and it was working out at 14k per day with 12 threads, or 12k with 6 threads.

I built it for family so it's not running that much now that I've handed it over, but it's here:

https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=3728342

One thing that Is really like to see the CPUs or Windows (not sure which manages clock speed these days) do better is manage the clock speed when running low priority tasks like BOINC projects. Ideally they would run at the base clock speed and then the CPU would still be able to turbo for other higher priority tasks. It's fine if you've got great cooling, but my last build was a small desktop with the stock fan so it got quite loud under load. Anyone know if there's a way to make that happen?

What helps me quite often in hotter months is simply to set windows to a different power profile. This Ryzen 1800x will run all day at 3.2 ghz just fine vs. the 3.7 it runs stock with a lot less noise and heat generation. I know that doesn't help much but it doesn't seem to be too horrible hit on performance. I'm looking at putting together a mini desktop with that 3500x so that's very good to know.

Something I'm observing at 100% CPU load on this ryzen with Rosetta is that the CPU power is amazingly low. It's using 77 Watts whereas usually when it's pegged at 100% it sits around 113 - 119. Is that because Rosetta doesn't take advantage of SSE or AVX? I don't know enough about either apart from when the Ryzen and other processors are doing AVX tasks they tend to run hotter. I don't think I'm going into swap since the CPU stays at a consistent 100% and BoincTasks is reporting 97-98% CPU usage per thread.

I had far too much Asteroids cached on some machines so I didn't get much Rosetta work so it's hard to tell what I'm getting for rac.

I recently built a 3600X and it was working out at 14k per day with 12 threads, or 12k with 6 threads.

I built it for family so it's not running that much now that I've handed it over, but it's here:

https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=3728342

One thing that Is really like to see the CPUs or Windows (not sure which manages clock speed these days) do better is manage the clock speed when running low priority tasks like BOINC projects. Ideally they would run at the base clock speed and then the CPU would still be able to turbo for other higher priority tasks. It's fine if you've got great cooling, but my last build was a small desktop with the stock fan so it got quite loud under load. Anyone know if there's a way to make that happen?

I use eFMer's TThrottle (https://efmer.com/download-tthrottle/) to control temperature. If you work out what temperature the high/noisy fan speed kicks in then you can set TThrottle to just below that and it will throttle BOINC tasks to below that CPU temp. It is better than setting a percentage CPU time in BOINC Manager as ambient temperature will affect the CPU running temperature. I find that my GPU especially will have it's temperature graph rise and fall over the temperature rise of the daytime.

I've got tthrottle on some laptops but I find it can make video playback really choppy when throttling.

---