[USRP-users] rx_samples_to_file issue
marcus.mueller at ettus.com
Mon Oct 6 17:36:12 EDT 2014
nice discussion indeed :)!
Phew, it really seems you got the most out of your system; if you want a
quick list of things in my head you'd want to check:
- NIC interrupt coalescing
- PCIe switching load: possibly we're just maxing out the PCIe
controller; sadly, I must admit I have no idea how likely that is, and
how to check.
So, to my understanding, linux kernels since some time now basically do
multi-threading in the sense that if you have a multicore system, IO
handling, especially flushing buffers to disk, should occur concurrently
and transparently to your UHD application, and that's the point I was
after: Why do home-brew IO buffering with multiple threads, if your OS
just does the same, but better?
The strange thing here is, and that indicates that even user land
multithreading might not be able to fix this, is that the GNU Radio
example doesn't affect the number of overruns -- it should, because
modern GNU Radio (with the default thread-per-block scheduler) is
inherently multithreaded. Stream-to-vector doesn't actually do
something, so it *shouldn't* help, but just "redefines" the stream item
size. Maybe the magic here lies in the fact that it's an additional
block, adding another layer of buffers.
So, I'll give Robert's mail a few thoughts and head over to bed,
On 06.10.2014 21:48, Peter Witkowski via USRP-users wrote:
> Forgot to add:
> I have all the recommended kernel tweaks for 10GigE running and found no
> difference in number of overflows vs. the network buffer size once you pass
> the point where UHD no longer throws a warning for your network buffer size
> being too small (if I recall correctly I think this is at 32MB or so?). I
> can provide a copy of my sysctl.conf file if necessary.
> I really hope that there's a kernel setting that I'm not using properly,
> otherwise I just don't see how the single-threaded approach can work for
> high data rates (in both the provided UHD programs and in user-built
> GNURadio programs). As I have mentioned, it is possible to get GNURadio to
> do some multi-threading and concurrent producer/consumer behavior, but this
> requires a re-purposing of the stream-to-vector block as it currently
> stands (or "rolling your own").
> On Mon, Oct 6, 2014 at 3:33 PM, Peter Witkowski <pwitkowski at gmail.com>
>> Great discussion so far.
>> I think my point is that earlier I saw that it was alluded to that if
>> rx_samples_to_file fails, then your disk subsystem is to blame and you
>> should invest in more hardware. My point was that the application in
>> general can benefit from multi-threading and internal buffering, otherwise
>> overflows are a fact of life for high data rates. While the OS can do more
>> buffering, you are required to do periodic (at a rather high rate) reads of
>> the USRP, otherwise its buffers will overflow. This is why it is critical
>> to minimize the time between two successive reads via threading and having
>> a buffer that can concurrently be added onto (from the device) and
>> processed (by writing to disk).
>> I think we can agree that there's a delta between what the I/O can sustain
>> (and what is displayed on a synthetic benchmark) versus what is actually
>> happening on the machine. There's potentially a good deal of CPU cycles
>> that occur between two reads (a vast percentage of these are during disk
>> I/O, especially if the device is "busy") that seem to be causing the
>> overflow. Stated differently, I can have a bunch of blocking reads occur
>> and sustain the needed rate. Similarly (and this is what I see via my
>> benchmarks), I can have a series of blocking writes occur and sustain the
>> needed data rate. However, the combination thereof (as well as the fact
>> that the kernel likes to preempt things) might be unsustainable in a single
>> threaded application. That is, by the time I get to the next read, enough
>> time has passed to overwhelm the USRP read buffers.
>> My point with the dirty background ratio is that when set to zero, I have
>> a predictable amount of time that I block for each write. My understanding
>> is that once triggered, background flushes will attempt to write at 100%
>> device speed. During this time, additional device access (such as those
>> being called by the application) will be blocked (sequential writes are
>> therefore performance degraded). When set to 0, I see in iotop that my
>> writes are a consistent value which is right about where resource monitor
>> shows the incoming 10GigE bandwidth to be. Setting the
>> dirty_background_ratio any higher and the "Total Disk Write" and "Actual
>> Disk Write" differ wildly. "Actual Disk Write" will spike at times, and
>> these spikes correlate with additional overflow errors. When the device is
>> more stable (i.e. "Actual Disk Write" is consistent) I see a huge reduction
>> in overflows. Stated differently, I'm OK with a flush, as long as its a
>> quick one. In the cases of larger flushes (as is the case with a
>> dirty_background_ratio greater than 0), I can almost guarantee an overflow
>> (I can't get around the mainloop fast enough if the device is busy with a
>> long write). Note that for my buffered application, I do run a default
>> background ratio.
>> The only thing that I got to work was buffering, and the easiest way I
>> know to accomplish this is to place a stream-to-vector block (in GNURadio)
>> between the USRP source and the file sink. The vectorization effectively
>> works as a buffer, and was the only way I was able to get 100 MS/s working.
>> Other things I tried:
>> 1. Using CPU affinity and shielding.
>> 2. Increasing processes priority to 99 and SCHED_FIFO.
>> Also, I can confirm that rx_samples_to_file and uhd_rx_cfile (as well as
>> GNU Radio application that does effectively the same thing, but controls
>> CPU affinity) all behave the exact same way in terms of the number of
>> overflows I have encountered.
>> If there are any other kernel tweaks that I can try, please let me know.
>> However, in practice, my CPU is not fast enough to get through the mainloop
>> of rx_samples_to_file quickly enough. As discussed above, buffering was
>> the only way I was able to consistently ensure that my data was saved and I
>> didn't encounter an overflow.
>> If you would like, I can provide whatever benchmarks of the system that
>> you would like. Perhaps I am not using the correct tool when I quote my
>> minimum write speeds to disk.
>> On Sat, Oct 4, 2014 at 9:14 AM, Marcus Müller <usrp-users at lists.ettus.com>
>>> Hi Peter,
>>> didn't mean to confuse you! Actually, my job is doing the opposite (ie.
>>> providing useful information), and thus let me just shortly follow up on
>>> On 03.10.2014 17:44, Peter Witkowski via USRP-users wrote:
>>> So I'm confused.
>>> You state that if I can't use rx_samples_to_file, my system is failing to
>>> perform as specified to write data out, then you give an example of several
>>> things that can happen to create a stochastic write speed (which I totally
>>> understand and agree with). Given that writes can be stochastic, why is
>>> there not a software buffer implemented in the UHD sample code to account
>>> for such issues?
>>> Well, because that's, in my opinion, an operating system's job. Being a
>>> code example, rx_samples_to_file just *musn't* contain the complexity
>>> introduced when you try to implement buffering functionality smarter than
>>> what your OS can do. And, I do think it's nearly impossible to be smarter
>>> than the linux kernel when optimizing writes -- *but* you'll have to tell
>>> your kernel what you want, as a user. The kernel, as it is configured by
>>> any modern distribution by default, won't do enormous write buffers,
>>> because that's not what the user usually wants, increasing the risk of data
>>> loss in case of system failure, and because you usually don't want to spend
>>> all of your RAM on filesystem buffers. In your 64GB RAM case, though,
>>> default buffer sizes should suffice, I guess, so I'm a bit out of clues
>>> It is definitely not very hard to increase these buffers' sizes, so I
>>> encourage you to try it and see if that solves your problem. Now, I must
>>> admit that up to here I was always assuming you hadn't already played
>>> around with these values, if this is not the case, please accept my
>>> I understand that it's meant to be an example, but I've
>>> also seen it referenced as being used effectively as a debugger or test for
>>> people having issues (i.e. recommendation to use the UHD programs in place
>>> of GNURadio to resolve issues).
>>> ...and it's done many users and thus Ettus a great job of supplying
>>> basic functionality! The fact that it works in almost any situation with
>>> this very minimalistic approach (repeated recv->write) proves that UHD is
>>> in fact a rather nice driver interface, IMHO. The fact that GNU Radio
>>> sometimes solves issues that rx_samples_to_file can't indicates exactly the
>>> buffering approach to be helpful. But in that case, buffering is not
>>> increased by increasing kernel buffer sizes, but by introducing GNU Radio
>>> buffers between blocks. The USRP source (Martin, scold me if I say
>>> something stupid) is not really much smarter than rx_samples_to_file: It
>>> recv()s a packet of samples, and returns these samples from the work
>>> function, and then GNU Radio takes care of shuffling and buffering that
>>> data. Basically, GNU Radio behaves much like an operating system from the
>>> source block's point of view.
>>> Also, in terms of benchmarking, I'm quoting minimum values, not averages.
>>> I agree with you that average values are pointless, and in reality the disk
>>> subsystem needs to perform when called up. My minimum values for a 4 disk
>>> RAID0 with a dedicated controller are well within the data rate that I am
>>> Well, I'll kind of disagree with you: If your minimum write rate of your
>>> system was bigger than the rate rx_samples_to_file causes, then you
>>> wouldn't see the problem. The point, I believe, here is that the storage
>>> system does not only consist of the hardware side of your RAID, but also on
>>> your complete operating environment. Something slows down how fast data is
>>> written to the RAID.
>>> I think we both would expect the following to happen:
>>> (blocks until a packet of samples has arrived. Instantly returns if
>>> it has before the call)
>>> write(file_handle, recv_buff)
>>> (instantly returns, because writing should hit a buffer that the
>>> operating system transparently pushes out to a disk. If buffer is full,
>>> then block until enough space in buffer -- unless your filesystem is
>>> mounted with some sync option...)
>>> Now, if your RAID is definitely fast enough, the write buffer should
>>> never get full. My hypothesis here is that either, your buffer size is just
>>> to small, and a block of samples doesn't fit and has to be written out
>>> instantly (which is unlikely), or something else occupies your system. That
>>> might be just the fact that 400MB/s (are we talking about an X3x0?)
>>> inevitably places a heavy load on things like PCIe busses and CPUs, and
>>> that introduces a bottleneck in your storage chain which isn't there if you
>>> "just" benchmark without the USRP. Also, the rather smallish sizes of
>>> network packets dictate that journalling file systems introduce a very bad
>>> overhead -- I don't know if you benchmarked with files on a journaling file
>>> system and a (network packet size - header) block size...
>>> Is there an example system that can handle sustained data capture from the
>>> USRP at (or near the limits) of 10GigE or the PCIe interfaces (maybe the
>>> requirement is enterprise class PCIe SSDs)? I'm running a two socket Xenon
>>> system (two hex core processors) with 64GB of RAM. How much more hardware
>>> should I throw at the problem to be able to sample/write at 100MS (half of
>>> what is quoted on the website for bandwidth for the 10GigE kit) using the
>>> provided code?
>>> Definitely a nice system! I must admit that I don't have access to a
>>> comparable setup, and thus I can't really offer you any first-hand
>>> experience. Maybe others can.
>>> I think the issue here is that the code itself can't simply get through
>>> it's main loop fast enough. There's a difference between data bandwidth
>>> and CPU throughput. The sequential nature of the code means that if any
>>> weird stuff happens (your example was a good set of kernel related hilarity
>>> that can lead to stochastic timing) you will have overflows since you
>>> cannot read fast enough. This is why a 90% solution for my application was
>>> to just set the dirty_background_ratio to 0 and also why redirection to
>>> /dev/null makes overflows go away.
>>> This is interesting, as dirty_background_ratio is the percentage at
>>> which the kernel should start writing out dirty pages in the background.
>>> Now, I'm the one who's confused, because I would have expected this to
>>> negatively impact performance. On the other hand, 0 (at least in my head)
>>> does not make very much sense, maybe it's semantically identical to 100%?
>>> Are you swapping (64GB would tell me you shouldn't have swap or extremly
>>> low swappiness)?
>>> On the other hand, it might really be that storage is not the bottleneck
>>> here, and in fact maybe the CPU gets saturated. Now, you said that writing
>>> to /dev/null solves your problem. Do your RAID or filesystem consume a lot
>>> of CPU cycles? This is an interesting mystery...
>>> With either method I didn't have to
>>> wait for a large write cache to flush before moving on to the next read
>>> from the USRP. Note that there can also be things that happen on the read
>>> side as well. Does this mean that I can only run the code on an RTOS?
>>> No :) UHD has it's own incoming buffer handlers, but as you already
>>> said, in this high performance scenario, you might be totally right, and
>>> our single-threaded approach just doesn't cut it. Maybe dropping in some
>>> asynchronous storage IO would help -- but I hate seeing that blowing up in
>>> example users' faces, so I guess the fact that it doesn't work with a
>>> system as potent as yours with the sample rates as high as you demand might
>>> actually be a shortcoming of the examples that isn't going to be fixed.
>>> As a final note, my understanding is that GNURadio and the USRP were
>>> developed for domain experts in DSP to use.
>>> These are SDR frameworks and devices, respectively. The idea is to offer
>>> people with the opportunity to build awesome DSP systems using universally
>>> usable SDR blocks (GNU Radio) and universal software radio peripherals, so
>>> well, they certainly address DSP people, but they shouldn't be hard to use.
>>> These users may or may not
>>> have prior experience in software. As a result, I'd recommend perhaps
>>> adding a buffered example or have the USRP GNURadio block allow for
>>> That is something we might consider. On the other hand, when someone
>>> goes as far as you do, maybe having an example that does the buffering in a
>>> separate thread (or even process) isn't worth that much -- in the end, one
>>> will want to write one's own high performance application, and that will
>>> include handling such data rates.
>>> Otherwise, I just don't see how you can advertise 200 MS/s
>>> (maybe even a simple "buffer" block in GNURadio would do the trick?).
>>> Well, the devices support these rates, and our driver is able to
>>> withstand these rates and sustain them without hitting CPU barriers due to
>>> having too much overhead. That's awesome (ok, I might be biased, but *I*
>>> think it's awesome). I don't feel ashamed because on your specific setup,
>>> we can't find a way to make any of our generic examples deliver the full
>>> rate of rx streams to storage -- we sell RF hardware, and not storage
>>> infrastructure, and the point of the examples is demonstrating the usage of
>>> UHD, and not holding a lecture on high performance storage handling. I
>>> wish, though, that we could solve your problem.
>>> Now, GNU Radio/gr-uhd does in fact come with an application called
>>> uhd_rx_cfile, which is more or less a clone of rx_samples_to_file using
>>> gr-uhd and GNU Radio instead of raw UHD. Does that work out for you?
>>> understand that this is theoretical limit of the bus, but if there doesn't
>>> exist a driver or other software to make use of this, the practical limit
>>> becomes much, much smaller.
>>> Well, UHD seems to be able to sustain these rates, if you write to
>>> /dev/null, right? So the practical limit for UHD is definitely not being
>>> I have another --maybe even practical-- suggestion to make: Roll your own
>>> mkfifo /tmp/mybuffer #assuming tmpfs is in ram
>>> dd if=/tmp/mybuffer of=/mount/raid_volume/data.dat & #start in
>>> background; you could play around with block sizes using the bs= option of
>>> rx_samples_to_file --file /tmp/mybuffer [all the other options]
>>> By the way: Thanks for bringing this up! We know that recording samples
>>> is a core concern of many users.
>>>  https://www.kernel.org/doc/Documentation/sysctl/vm.txt
>>> On Fri, Oct 3, 2014 at 10:55 AM, Marcus Müller <usrp-users at lists.ettus.com> <usrp-users at lists.ettus.com>
>>> I have to agree with Marcus on this. Also, keep in mind that storage is
>>> really what an operating system should take care of in any "general
>>> purpose" scenario, ie. that as long as I just write to a file, I'd expect
>>> that the thing in charge of storage (my kernel / the filesystems / block
>>> device drivers) does the best it can to keep up. If I find myself in a
>>> situation where my specific storage needs dictate a huge write buffer,
>>> changing the application might be one way, but as I'm responsible for my
>>> won storage subsystem, I could just as well increase the cache buffer
>>> sizes, and let the operating system handle storage operation. If your RAID
>>> is really performing as well as it is benchmarked to, then this should not
>>> be one of your problems. All rx_samples_to_file does is really sequentially
>>> writing out data at a constant rate, which is the most basic write
>>> benchmark I can think of.
>>> If your storage subsystem (filesystem + storage abstraction + raid driver
>>> + interface driver + hard drive interface + hard drives + hardware caches)
>>> can't keep up, it's failing to perform as specified, simple as that. In
>>> this case, saying that the application needs to be smarter when dealing
>>> with storage seems like a bit of a cop-out to me ;)
>>> I'd like to point out that most benchmarks use heavily averaged numbers
>>> for write speeds etc. UHD on the other hand kind of demands soft real-time
>>> performance of a write subsystem, which is a lot harder to fulfill. This
>>> comes up rather frequently, but I have to stress it: you need a fast
>>> guaranteed write rate, not only an average one, and as soon as your
>>> operating system has to postpone writing data, it has to have enough
>>> performance to catch up whilst still meeting continued demand. This is
>>> general purpose hardware running general purpose OS with dozens of
>>> processes, and you can't just say "every single component is up to my task,
>>> thus my system suffices", because everything potentially blocks everything!
>>>  e.g. because the filesystems needs to calculate checksums, update
>>> tables, another process gets scheduled, a device blocks your PCIe bus, your
>>> platters randomly need a bit longer to seek, you reach the physical end of
>>> an LVM volume and have to move across a disk, an interrupt does what an
>>> interrupt does, some process is getting noticed on a changing file
>>> descriptor, DBUS is happening in the kernel, token ring has run out of
>>> tokens, thermal throttling, bitflips on SATA leading to retransmission,
>>> some page getting fetched from swap...
>>> On 03.10.2014 15:34, Marcus D. Leech via USRP-users wrote:
>>> One has to keep firmly in mind that programs like rx_samples_to_file are
>>> *examples* that show how to use
>>> the underlying UHD API. They are not necessarily optimized for all
>>> situations, and indeed, one could
>>> restructure rx_samples_to_file to decouple UHD I/O from filesystem I/O,
>>> using a large buffer between them.
>>> The fact is that dynamic performance of high-speed, real-time, flows is
>>> something that almost-invariably needs
>>> tweaking for any particular situation. There's no way for an example
>>> application to meet all those requirements.
>>> But the fact also remains that for *some* systems, rx_samples_to_file
>>> (and uhd_rx_cfile on the Gnu Radio side)
>>> are able to stream high-speed data just fine as-is.
>>> On 2014-10-03 09:26, Peter Witkowski via USRP-users wrote:
>>> To say that the issue is just because the disk subsystem can't keep up is a bit of cop-out.
>>> I had issues writing to disk when the incoming stream was 400MB/s and my RAID0 system was benchmarked at being much higher than that.
>>> The issue that I've been seeing stems from the fact that it appears that you cannot concurrently read/write from the data stream as its coming in. In effect you have a main loop that reads from the device and then immediately tries to write that buffer to file. If you do not complete these operations in a timely fashion overflows occur.
>>> One way to solve (or at least band aid the issue) is to set your dirty_background_ratio to 0. I was able to get writing to disk working somewhat with this setting as it is more predictable to directly write to disk instead of having your write cache fill up and then having a large amount of data to push to disk. That said, my RAID0 array is capable of such speeds and even then I was getting a few (but much reduced) overflows.
>>> The one surefire way I know of getting this working (even on a slow disk system) is to buffer the data. The buffer can then be consumed by the disk writing process while being concurrently added onto by the device reader. The easiest way to test buffering (that I've found) is to simply set up a GNURadio Companion program with a stream-to-vector block between the USRP and file sink blocks. This is exactly what I am doing currently since even with a very powerful system, I could not get data saved to disk quickly enough given the aforementioned issues with the provided UHD software.
>>> On Thu, Oct 2, 2014 at 11:48 PM, gsmandvoip via USRP-users <usrp-users at lists.ettus.com> <usrp-users at lists.ettus.com> <usrp-users at lists.ettus.com> <usrp-users at lists.ettus.com> wrote:
>>> Thanks Marcus for your replies. Yes O gone away.
>>> On Thu, Oct 2, 2014 at 5:50 PM, Marcus D. Leech <mleech at ripnet.com> <mleech at ripnet.com> <mleech at ripnet.com> <mleech at ripnet.com> wrote:
>>> with rx_samples_to_file without _4rx.rbf, Initially I tried on my i3, 4GB ram, it gave me
>>> some OOOO but was lesser than earlier, but I do not understand, my most of the ram capacity and processor was sitting idle while it shows OOOO, why is this strange behaviour The default format for uhd_rx_cfile is complex-float, thus doubling the amount of data written compared to rx_samples_to_file.
>>> You can't just use CPU usage as an indicator of loading--if you're writing to disk, the disk subsystem may be much slower than you think, so the
>>> "rate limiting step" is writes to the disk, not computational elements.
>>> Try using /dev/null as the file that you write to. If the 'O' go away, even at higher sampling rates, then it's your disk subsystem.
>>> using uhd_rx_cfile getting similar result, but strangely, why it is low, at 4M sampling rate it was higher???
>>> On Thu, Oct 2, 2014 at 9:27 AM, Marcus D. Leech <mleech at ripnet.com> <mleech at ripnet.com> <mleech at ripnet.com> <mleech at ripnet.com> wrote:
>>> On 10/01/2014 11:46 PM, gsmandvoip wrote:
>>> Yes I am running single channel, but when trying to achieve my desired sampling rate without _4rx.rbf, it says, requested sampling rate is not valid, adjusting to some 3.9M or so. sorry for misleading info I gave earlier, I have i3, with 32 bit and i7 with 64 bit, but getting same result on both machines
>>> Here is my command to capture signal:
>>> ./rx_samples_to_file --args="fpga=usrp1_fpga_4rx.rbf, subdev=DBSRX" --freq "$FC" --rate="$SR" $FILE --nsamps "$NSAMPLES"
>>> and here is its output:
>>> Creating the usrp device with: fpga=usrp1_fpga_4rx.rbf, subdev=DBSRX...
>>> -- Loading firmware image: /usr/share/uhd/images/usrp1_fw.ihx... done
>>> -- Opening a USRP1 device...
>>> -- Loading FPGA image: /usr/share/uhd/images/usrp1_
>>> fpga_4rx.rbf... done
>>> -- Using FPGA clock rate of 52.000000MHz...
>>> ERROR: LOOKUPERROR: INDEXERROR: MULTI_USRP::GET_TX_SUBDEV_SPEC(0) FAILED TO MAKE DEFAULT SPEC - VALUEERROR: THE SUBDEVICE SPECIFICATION "A:0" IS TOO LONG.
>>> The user specified 1 channels, but there are only 0 tx dsps on mboard 0.
>>> Don't use the _4rx image if you don't need it.
>>> The USRP1 only does strict-integer resampling, and with a master clock (NON STANDARD FOR USRP1) of 52.000MHz, 4Msps is not a sample rate
>>> that it can produce. Try 5.2Msps or 4.3333Msps.
>>> At 5.2Msps, it's recording at roughly 20.8Mbytes/second, so your system needs to be able to sustain that for at least as long as the capture lasts.
>>> USRP-users mailing listUSRP-users at lists.ettus.comhttp://lists.ettus.com/mailman/listinfo/usrp-users_lists.ettus.com
>>> USRP-users mailing listUSRP-users at lists.ettus.comhttp://lists.ettus.com/mailman/listinfo/usrp-users_lists.ettus.com
>>> USRP-users mailing listUSRP-users at lists.ettus.comhttp://lists.ettus.com/mailman/listinfo/usrp-users_lists.ettus.com
>>> USRP-users mailing list
>>> USRP-users at lists.ettus.com
>> Peter Witkowski
>> pwitkowski at gmail.com
> USRP-users mailing list
> USRP-users at lists.ettus.com
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the USRP-users