89 replies to this topic

[WanWizard]

Afaik nothing has changed to this in quite some time. For modern 4K sector disks, alignment is indeed important.

 

If I check here, small(ish) devices like my 32Gb USB stick, start at 1049Kb, but my disk starts at 4096B.

 

1049Kb strikes me as a bit odd, it means the first Mb is gone (as it is in Kb and not ? 

[anudanan]

I´ve found some Artikel which says 1049kB ist normal, because an 4K allignment ist for example 1024*1024 = 1048576 Byte (first mbyte of the drive)

 

1049kB means 1049000 Byte nearly the same as 1048576. I don´t not why the now you decimal writings.

[anudanan]

now I will make some test if clustersize of 2MByte brings more performance than 256Kbyte for parallel recordings and cutting, moving

[Erik Slagter]

It will surely make a difference but not a BIG difference, maybe not even noticable.

[anudanan]

I made 12 recordings, one cutting job and one moving Job

 

I see that during my tests the I/O load is ok for some minutes  and between 50-60 percent during reading and writing from/to the drive The write rate is 15-30 MByte/s. But sometimes I see a very high I/O load 90-96 percent and only 1-2 Mbyte/s write and 1-2 Mbyte/s read

 

Why ist the I/O load so high during the low I/O rates to/from the drive. 

 

On my older et8000 with openpli 4.0 that was not happend during those tests. I think the uno4kse is from the hardware view in the same class then the et8000 and must be a little bit faster. But what is the reason for the sometimes bad performance and high I/O load.

 

When I make my simulated tests with dd for recording 8 files, cutting and moving files the load is allways better during write nearly

[anudanan]

Now I´ve found out a very interesting result

 

I´ve startet only a movie cuttingjob  with the same movie on my unx4kse with openpli 4.0 and at the same time on my et9200 mit openpli 4.0. On both boxes with  the internal 2TB drive inside the boxes.

 

The uno4kse has a modern BCM7252S 2x1,7 Ghz ARM core  and the ET9200 has an old BCM 7405 (I think it is an 400Mhz dualcore MIPS)

 

The result is, the cutting Job on both boxes need nearly the same time, but the I/O load on the old BCM7405 was  between 6% and 16%.. On the modern ARM BCM7252S the I/O load during this cutting Job was mostly high above 50%-96%

 

 

This is strange. I this the modern ARM BCM and the old MIPS BCM have both DMA for the SATA interface and other periphals. But I think the driver für die ARM ist poor and don´t use DMA efficently

 

Do you think that is possible? Is that problem than on all ARM boxes because they all use the same drive of broadcom. So all ARM boxes of VU+, Dream, and other have the same poor performance by intensive disk I/O

 

I was very surprised about that because I thought that the uno4kSE box must have the same or higer performance as my old ET8000. The old box had no problems with recordings, cuttings and moving movies in parallel. But the new one has it and I thing that is a software and not a hardware problem

 

Do you think I´m right?

 

Do you think there will come newer and better drivers in the future for solving that problem?

 

I think today about a change from ARM to MIPS box but I can´t find a modern receiver box with  FBC tuner and MIPS to an acceptablke price. 

 

I can´t use my ET8000 because that box has an ethernetproblem /link down/up) and I have gave them back on guarantee and bought  the uno4kse then.

[anudanan]

Is it possible to switch the driver operation mode from PIO to DMA for die ARM cores? I think the driver works today in PIO mode and die Core must receive and sent the data by himself and that maybe the reason for the high I/O load of the CPU

[anudanan]

Is it possible to get an openpli version 4.0 with a kernel 3.14 for uno4kse to test the hdd performance ? I saw for  uno4k it exists an openpli 4.0 and the box is very similar to the uno4kse. For tests it is ok if the lcd display doesn´t work

[WanWizard]

I think there is no BSP for OpenPLi 4 for the Uno 4K SE.

 

You might be able to backport providing that the current kernel is also 3.14. If not, all stops, as the drivers dictate the kernel being used.

 

If you're able, you need a linux build environment with gcc 4.x, checkout the 4-release branch, and build an image.

[anudanan]

I think I will check the performance with openATV 6.2 (rocko based). I don´t know exactly which kernel they use I think the kernel is newer

 

VTI use the same as openpli 6.1 (Kernel 4.1.20), I´ve made the test and they have the same problems

 

Hopefully the kernel with drivers in openatv 6.2 have more performance and not so much CPU time in I/O section

I don´t know which kernel dream multimedia use for their ARM boxes

 

If the BCM ARM Technology is the future for the boxes (I think so) then it is important to found the reason for the poor hdd performance situation which Ive seen

[MiLo]

Is it possible to switch the driver operation mode from PIO to DMA for die ARM cores?...

That's PC terminology. There's no such thing as "PIO mode" for these chips.

[MiLo]

... but the I/O load on the old BCM7405 was  between 6% and 16%.. On the modern ARM BCM7252S the I/O load during this cutting Job was mostly high above 50%-96% ...

How are you measuring this?

[anudanan]

So you mean the BCM chips make allways DMA transfer to the hdd drives?

 

But what is the reason for the very high CPU load for I/O with the BCM7252 and why has the older BCM 7405 a massiv lower I/O load

 

Maybe a driver bug, but how is the way to fix it. 

[anudanan]

I use the Linux top command

[anudanan]

Before testing images from openatv I will make some tests with  a patched version of MovieCut which starts mcut with a nice lever -n+19 to give the cutting job a more lower priority than enigma. Additional I replaced the "mv" command in EMC also with "nice -n+19 mv" to see, if I can protect the recordings for interrupts.

Edited by anudanan, 9 January 2018 - 14:11.

[Erik Slagter]

Nice won't change a thing. We're talking I/O bound processes here, they take almost no CPU slices.

[Erik Slagter]

So you mean the BCM chips make allways DMA transfer to the hdd drives?

 

But what is the reason for the very high CPU load for I/O with the BCM7252 and why has the older BCM 7405 a massiv lower I/O load

 

Maybe a driver bug, but how is the way to fix it. 

What do you count as "high cpu load"? How do you determine the CPU load?

 

Some receivers apparently are using polling for some DVB functions, so they're CPU load is also way above "low".

[anudanan]

I see the different CPU loads with the top tool in teh ET9200 and uno4kse.

I see that the old ET9200 has lower load than the uno4kse by doing the same thing. My question than is why is the load on the uno4kse higher than on the ET9200.

 

I have the idea that the cpu I/O load for the data to/from disk of mv-process and mcut-process goes lower if I use a higher  nice value for that processes.

[Erik Slagter]

You should at least compare the load when doing disk access compared the load when idle.

 

Nice value is relative. It means the specific process gets a lesser share of the total available cpu time. As long as the load is lower than 100 * (# of cpu's), there are cpu time slices unused. Unused cpu slices are again distributed against over all processes. Until no process are left that are waiting for the processor ("runnable").

 

So if you're only having one process that is actually using the CPU, you can nice it however you like, it will NOT make a difference. It will only make a difference if there is at least one OTHER process that is continuously running (not waiting for I/O). Now that other process gets most of the cpu slices, your first processes gets considerately less.

 

But as enigma is using almost no CPU time slices, this situation is academic here.

 

Just like people that do not properly understand technical operating system basics, should not judge memory usage, they should also not judge cpu usage.

Edited by Erik Slagter, 9 January 2018 - 19:43.

[anudanan]

The cpu load for I/O on the un4kse ist during mcut execution for ony one moviecut between 50 and 80 percent and the rest goes over sys, usr and a bit for idle. On my et9200, which makes the same mcut execution for the same movie in the same real time, the CPU I/O load is between 8-16 percent and has more idle free.

 

It was similar by my ET8000 and the ET8000 make no interrupts in records when in parallel mcut and mv runs.

 

 

That is the reason why I think the CPU I/O load is to high on uno4kse

 

Du you think everything is ok with the load on uno4kse?