Finding Out the Bottleneck of the PC

[GreatEmerald]

I'm trying to determine what is the limiting factor in my PC build (it's already slightly dated, and I'm considering an upgrade of one of the parts). Since I use my PC for recording high quality screen captures (that's when I feel that the PC is acting rather slow), I decided to monitor the performance of the CPU and the disk during recording in order to determine which one is more likely a limiting factor.

Here are my results:
CPU cores 0 and 1
CPU cores 2 and 3
Disk write in SI MB/s

First half of the graph shows the performance while recording a 3D game in full HD using Fraps, and the latter half shows CrystalMark disk benchmark being run.

The drops in disk write speed correspond to lag spikes I get in the game, as well as to when Fraps reaches the 4 GiB file limit and decides to split the video.

Now, how do I interpret the data? CPU usage doesn't seem to be at its height (especially for core 1), although it's still rather high, given that on multi-core setups, the cores still have to wait for the others, so the core usage is not always at its peak.

The disk data is even more puzzling. CrystalDiskMark has 4*2 stages, 4 of which are reflected here (as only the write data is displayed here): sequential write with 1024KiB blocks, random write with 512KiB blocks, random write with 4KiB blocks and random write with 4 KiB blocks and 32 Queue Depth. In my graphs, the first block after I stopped recording was when the program was preparing the data (4000 MB), so it wasn't tested yet. The test phases are from 12:28:00 to 12:29:30, 12:29:30 to 12:30:30, 12:30:30 to 12:31:00 and 12:31:00 to 12:31:30, if I recall correctly.

Why are the speeds so different? Why is a sequential write so much faster than a random write? And how does it compare to the real-life data - is my disk during the recording at its maximum capability for the type of data being written to disk, or is it not even close to the limit?

Lastly, why do I get lag spikes when Fraps changes the file? From the CPU graphs, you can see that CPU usage skyrockets at that point, then falls down rapidly, and then resumes, while disk write halts for a while, and then resumes. What causes the lag, then - the disk, or the CPU?

And, most importantly, which one is stressed more overall - the disk or the CPU?

[Fordy]

For this sort of use, and unless your CPU is properly ancient (which I assume it isn't from the four cores), upgrading to an SSD would see the biggest benefit.

Random writes are always slower because the data is being written all over the place, not, sequentially. Even in an SSD where heads and spindles don't need to physically move to change where the data is written, they still need to locate the next target and redirect the incoming bits.

[GreatEmerald]

Yeap, it's a Phenom II X4 925, so it should be rather good.
I did a bit more research about this, and it appears that the spikes are due to the way Fraps handles things, so it's likely an audio issue... The disk, WD Caviar Green (I launch the game from another disk, so the OS and the game don't interfere with the recordings), should be enough for what is written to it, too, as large video files are sequential, from what I know. At least given low amounts of fragmentation.

Though now I wonder how that scales when recording different games. I would assume that the overhead that is created by the recording software is more or less static - data is written on disk at a fairly constant rate (unless I am recording a completely static scene).

What puzzles me is how some more demanding games, such as Unreal Tournament 3 or Mass Effect 3, tend to lag, even if not recording, even though according to the data, my CPU core (since they are single-threaded) usage is always around 60%, and same thing goes for the graphics card (the maximum load on the GPU was 80%). Though on the CPU, cores other than the one the game was on also showed an increase in usage, but I guess it's just because other processes were attributed to those cores. So if everything uses only 80% of resources on maximum load, why do I get drops in framerate?

My GPU is an AMD Radeon HD 4890, so it's also fairly capable.

[GreatEmerald]

Hm, well this is interesting. I found a place which gives me a reliably low framerate (Mass Effect 3, Citadel refugee housing), and monitored my resources for a while. Apparently my framerate goes low when there are many entities on screen - when looking over the Citadel with a lot of independently animating people, I get slowdowns like that. Interestingly enough, during that time, my GPU usage doesn't increase, but often quite the contrary - it decreases. However, I see that my CPU usage does indeed increase reliably, but also in an interesting fashion.

Here are the charts I made this time (running Mass Effect 3, 1080p resolution, not recording this time):
All CPU values combined
CPU core 3
CPU core 0
CPU core 2
CPU core 1

I got drops in framerate from 09:14:10 to 09:14:40, from 09:15:50 to 09:17:00 and from 19:17:40 to 19:18:00 - that is when I set the camera to overlook most of the people in the scene. Other times I set the camera to view a wall.

Each of these cores show some interesting things. Core 3 seems to change the most, closest to how the framerate and overall CPU usage changes. Core 0 seems to be fairly static and highly used. Core 2 usage is very low, although it seems to spike during framerate drops. Core 1 is similar to core 2, but its usage persists even after minimising the game.

That makes me wonder how the game utilises the CPU there... Since the usage is not very balanced, it seems that it's not using multiple cores very efficiently, but it still is using them. From what I gather, some parts of the game, like audio and animations, are sent to other threads, although the game still maintains a central thread for all of the game logic.

Just as a comparison, I used the first Unreal game, which is completely single-threaded, and told it to choke my CPU by using unaccelerated high definition shadows. This is how it looks before and after I set them on:
http://img560.imageshack.us/img560/8...rformance8.gif
Basically, it seems that the system is throwing around the thread across the cores (except for the one I highlighted there, which is oddly unaffected...). The end result is always 25%, but it shows the CPU struggling very clearly. So perhaps to some extent that could account for the spikes in cores 2 and 1 in Mass Effect 3...

And that makes me wonder: what is more important in this case, more cores or more performance per core? Unreal Engine 3 is somewhat multi-threaded, but as you can see, even that isn't that well balanced, so I would assume that more performance per core is more important in this case...