Re: [Qemu-devel] [RFC PATCH 0/6] virtio-trace: Support virtio-trace

[Amit Shah]

On (Tue) 24 Jul 2012 [11:36:57], Yoshihiro YUNOMAE wrote:
> Hi All,
> 
> The following patch set provides a low-overhead system for collecting kernel
> tracing data of guests by a host in a virtualization environment.
> 
> A guest OS generally shares some devices with other guests or a host, so
> reasons of any problems occurring in a guest may be from other guests or a 
> host.
> Then, to collect some tracing data of a number of guests and a host is needed
> when some problems occur in a virtualization environment. One of methods to
> realize that is to collect tracing data of guests in a host. To do this, 
> network
> is generally used. However, high load will be taken to applications on guests
> using network I/O because there are many network stack layers. Therefore,
> a communication method for collecting the data without using network is 
> needed.
> 
> We submitted a patch set of "IVRing", a ring-buffer driver constructed on
> Inter-VM shared memory (IVShmem), to LKML http://lwn.net/Articles/500304/ in
> this June. IVRing and the IVRing reader use POSIX shared memory each other
> without using network, so a low-overhead system for collecting guest tracing
> data is realized. However, this patch set has some problems as follows:
>  - use IVShmem instead of virtio
>  - create a new ring-buffer without using existing ring-buffer in kernel
>  - scalability
>    -- not support SMP environment
>    -- buffer size limitation
>    -- not support live migration (maybe difficult for realize this)
> 
> Therefore, we propose a new system "virtio-trace", which uses enhanced
> virtio-serial and existing ring-buffer of ftrace, for collecting guest kernel
> tracing data. In this system, there are 5 main components:
>  (1) Ring-buffer of ftrace in a guest
>      - When trace agent reads ring-buffer, a page is removed from ring-buffer.
>  (2) Trace agent in the guest
>      - Splice the page of ring-buffer to read_pipe using splice() without
>        memory copying. Then, the page is spliced from write_pipe to virtio
>        without memory copying.

I really like the splicing idea.

>  (3) Virtio-console driver in the guest
>      - Pass the page to virtio-ring
>  (4) Virtio-serial bus in QEMU
>      - Copy the page to kernel pipe
>  (5) Reader in the host
>      - Read guest tracing data via FIFO(named pipe) 

So will this be useful only if guest and host run the same kernel?

I'd like to see the host kernel not being used at all -- collect all
relevant info from the guest and send it out to qemu, where it can be
consumed directly by apps driving the tracing.

> ***Evaluation***
> When a host collects tracing data of a guest, the performance of using
> virtio-trace is compared with that of using native(just running ftrace),
> IVRing, and virtio-serial(normal method of read/write).

Why is tracing performance-sensitive?  i.e. why try to optimise this
at all?

> <environment>
> The overview of this evaluation is as follows:
>  (a) A guest on a KVM is prepared.
>      - The guest is dedicated one physical CPU as a virtual CPU(VCPU).
> 
>  (b) The guest starts to write tracing data to ring-buffer of ftrace.
>      - The probe points are all trace points of sched, timer, and kmem.
> 
>  (c) Writing trace data, dhrystone 2 in UNIX bench is executed as a benchmark
>      tool in the guest.
>      - Dhrystone 2 intends system performance by repeating integer arithmetic
>        as a score.
>      - Since higher score equals to better system performance, if the score
>        decrease based on bare environment, it indicates that any operation
>        disturbs the integer arithmetic. Then, we define the overhead of
>        transporting trace data is calculated as follows:
>               OVERHEAD = (1 - SCORE_OF_A_METHOD/NATIVE_SCORE) * 100.
> 
> The performance of each method is compared as follows:
>  [1] Native
>      - only recording trace data to ring-buffer on a guest
>  [2] Virtio-trace
>      - running a trace agent on a guest
>      - a reader on a host opens FIFO using cat command
>  [3] IVRing
>      - A SystemTap script in a guest records trace data to IVRing.
>        -- probe points are same as ftrace.
>  [4] Virtio-serial(normal)
>      - A reader(using cat) on a guest output trace data to a host using
>        standard output via virtio-serial.
> 
> Other information is as follows:
>  - host
>    kernel: 3.3.7-1 (Fedora16)
>    CPU: Intel Xeon address@hidden(12core)
>    Memory: 48GB
> 
>  - guest(only booting one guest)
>    kernel: 3.5.0-rc4+ (Fedora16)
>    CPU: 1VCPU(dedicated)
>    Memory: 1GB
> 
> <result>
> 3 patterns based on the bare environment were indicated as follows:
>                          Scores      overhead against [0] Native
>     [0] Native:          28807569.5               -
>     [1] Virtio-trace:    28685049.5             0.43%
>     [2] IVRing:          28418595.5             1.35%
>     [3] Virtio-serial:   13262258.7            53.96%
> 
> 
> ***Just enhancement ideas***
>  - Support for trace-cmd
>  - Support for 9pfs protocol
>  - Support for non-blocking mode in QEMU

There were patches long back (by me) to make chardevs non-blocking but
they didn't make it upstream.  Fedora carries them, if you want to try
out.  Though we want to converge on a reasonable solution that's
acceptable upstream as well.  Just that no one's working on it
currently.  Any help here will be appreciated.

>  - Make "vhost-serial"

I need to understand a) why it's perf-critical, and b) why should the
host be involved at all, to comment on these.

Thanks,

                Amit