1. I learned a bit more about the scheduler from Michael. With multiple CPUs, each CPU has a queue of processes that are either "on the CPU" (running) or waiting to be serviced (run) on that CPU. Those processes count as "migratable" in runqueue_t. Every now and then, the system checks all its run queues to see if a CPU is idle, and can thus "steal" (migrate) processes from a busy CPU. This is done in sched_balance().

   Such "stealing" (migration) has the positive effect that the process doesn't have to wait for getting serviced on the CPU it's currently waiting on. On the other side, migrating the process has effects on CPU's data and instruction caches, so switching CPUs shouldn't be taken too easy.

   If migration happens, then this should be done from the CPU with the most processes that are waiting for CPU time. In this calculation, not only the current number should be counted in, but a bit of the CPU's history is taken into account, so processes that just started on a CPU are not taken away again immediately. This is what is done with the help of the processes currently migratable (r_mcount) and also some "historic" average. This "historic" value is taken from the previous round in r_avgcount. More or less weight can be given to this, and it seems that the current number of migratable processes had too little weight over all to be considerend.

   What happens in effect is that a process is not taken from its CPU, left waiting there, with another CPU spinning idle. Which is exactly what I saw in the first place.

2. What I also learned from Michael was that there are a number of sysctl variables that can be used to influence the scheduler. Those are available under the "kern.sched" sysctl-tree:

   % sysctl -d kern.sched
   kern.sched.cacheht_time: Cache hotness time (in ticks)
   kern.sched.balance_period: Balance period (in ticks)
   kern.sched.min_catch: Minimal count of threads for catching
   kern.sched.timesoftints: Track CPU time for soft interrupts
   kern.sched.kpreempt_pri: Minimum priority to trigger kernel preemption
   kern.sched.upreempt_pri: Minimum priority to trigger user preemption
   kern.sched.rtts: Round-robin time quantum (in milliseconds)
   kern.sched.pri_min: Minimal POSIX real-time priority
   kern.sched.pri_max: Maximal POSIX real-time priority 

   The above text shows that much more can be written about the scheduler and its whereabouts, but this remains to be done by someone else (volunteers welcome!).

3. Now, while digging into this, I also learned that I'm not the first to discover this issue, and there is already another PR on this. I have opened PR kern/51615 but there is also kern/43561. Funny enough, the solution proposed there is about the same, though with a slightly different implementation. Still, *2 and <<1 are the same as are /2 and >>1, so no change there. And renaming variables for fun doesn't count anyways. ;) Last but not least, it's worth noting that this whole issue is not Xen-specific.

To test a different load on the system, I've started a "build.sh -j8" on a (VMware Fusion) VM with 4 CPUs on a Macbook Pro, and it nearly brought the machine to a halt - What I saw was lots of idle time on all CPUs though. I aborted the exercise to get some CPU cycles for me back. I blame the VM handling here, not the guest operating system.

I restarted the exercise with 2 CPUs in the same VM, and there I saw load distribution on both CPUs (not much wonder with -j8), but there was also quite some idle times in the 'make clean / install' phases that I'm not sure is normal. During the actual build phases I wasn't able to see idle time, though the system spent quite some time in the kernel (system). Example top(1) output:

    load averages:  9.01,  8.60,  7.15;               up 0+01:24:11      01:19:33
    67 processes: 7 runnable, 58 sleeping, 2 on CPU
    CPU0 states:  0.0% user, 55.4% nice, 44.6% system,  0.0% interrupt,  0.0% idle
    CPU1 states:  0.0% user, 69.3% nice, 30.7% system,  0.0% interrupt,  0.0% idle
    Memory: 311M Act, 99M Inact, 6736K Wired, 23M Exec, 322M File, 395M Free
    Swap: 1536M Total, 21M Used, 1516M Free
    
    PID USERNAME PRI NICE   SIZE   RES STATE      TIME   WCPU    CPU COMMAND
    27028 feyrer    20    5    62M   27M CPU/1      0:00  9.74%  0.93% cc1
      728 feyrer    85    0    78M 3808K select/1   1:03  0.73%  0.73% sshd
    23274 feyrer    21    5    36M   14M RUN/0      0:00 10.00%  0.49% cc1
    21634 feyrer    20    5    44M   20M RUN/0      0:00  7.00%  0.34% cc1
    24697 feyrer    77    5  7988K 2480K select/1   0:00  0.31%  0.15% nbmake
    24964 feyrer    74    5    11M 5496K select/1   0:00  0.44%  0.15% nbmake
    18221 feyrer    21    5    49M   15M RUN/0      0:00  2.00%  0.10% cc1
    14513 feyrer    20    5    43M   16M RUN/0      0:00  2.00%  0.10% cc1
      518 feyrer    43    0    15M 1764K CPU/0      0:02  0.00%  0.00% top
    20842 feyrer    21    5  6992K  340K RUN/0      0:00  0.00%  0.00% x86_64--netb
    16215 feyrer    21    5    28M  172K RUN/0      0:00  0.00%  0.00% cc1
     8922 feyrer    20    5    51M   14M RUN/0      0:00  0.00%  0.00% cc1 

    # dmesg | grep cpu
    vcpu0 at hypervisor0: Intel(R) Xeon(R) CPU E5-2680 v2 @ 2.80GHz, id 0x306e4
    vcpu1 at hypervisor0: Intel(R) Xeon(R) CPU E5-2680 v2 @ 2.80GHz, id 0x306e4

I was looking at top(1) to see that everything was running fine, and noticed funny WCPU and CPU values:

      PID USERNAME PRI NICE   SIZE   RES STATE      TIME   WCPU    CPU COMMAND
      2791 root      25    0  8816K  964K RUN/0     16:10 54.20% 54.20% myprog
      2845 root      26    0  8816K  964K RUN/0     17:10 47.90% 47.90% myprog

top's CPU state shows:

    load averages:  2.15,  2.07,  1.82;               up 0+00:45:19        18:00:55
    27 processes: 2 runnable, 23 sleeping, 2 on CPU
    CPU states: 50.0% user,  0.0% nice,  0.0% system,  0.0% interrupt, 50.0% idle
    Memory: 119M Act, 7940K Exec, 101M File, 3546M Free

    load averages:  2.14,  2.08,  1.83;               up 0+00:45:56        18:01:32
    27 processes: 4 runnable, 21 sleeping, 2 on CPU
    CPU0 states:  100% user,  0.0% nice,  0.0% system,  0.0% interrupt,  0.0% idle
    CPU1 states:  0.0% user,  0.0% nice,  0.0% system,  0.0% interrupt,  100% idle
    Memory: 119M Act, 7940K Exec, 101M File, 3546M Free

One option is to kick your operating system out of the window, but I still like NetBSD, so here's another solution: NetBSD allows to create "processor sets", assign CPU(s) to them and then assign processes to the processor sets. Let's have a look!

Processor sets are manipulated using the psrset(8) utility. By default all CPUs are in the same (system) processor set:

    # psrset
    system processor set 0: processor(s) 0 1

    # psrset -c
    1
    # psrset
    system processor set 0: processor(s) 0 1
    user processor set 1: empty

    # psrset -a 1 1
    # psrset
    system processor set 0: processor(s) 0
    user processor set 1: processor(s) 1

    # ps -u 
    USER  PID %CPU %MEM   VSZ  RSS TTY     STAT STARTED     TIME COMMAND
    root 2791 52.0  0.0  8816  964 pts/4   R+    5:28PM 22:57.80 myprog
    root 2845 50.0  0.0  8816  964 pts/2   R+    5:26PM 23:33.97 myprog
    #
    # psrset -b 0 2791
    # psrset -b 1 2845

    load averages:  2.02,  2.05,  1.94;               up 0+00:59:32        18:15:08
    27 processes: 1 runnable, 24 sleeping, 2 on CPU
    CPU0 states:  100% user,  0.0% nice,  0.0% system,  0.0% interrupt,  0.0% idle
    CPU1 states:  100% user,  0.0% nice,  0.0% system,  0.0% interrupt,  0.0% idle
    Memory: 119M Act, 7940K Exec, 101M File, 3546M Free
    Swap:

      PID USERNAME PRI NICE   SIZE   RES STATE      TIME   WCPU    CPU COMMAND
     2845 root      25    0  8816K  964K CPU/1     26:14   100%   100% myprog
     2791 root      25    0  8816K  964K RUN/0     25:40   100%   100% myprog

Now why this didn't happen by default is left as an exercise to the reader. Hints that may help:

    # uname -a
    NetBSD foo.eu-west-1.compute.internal 7.0 NetBSD 7.0 (XEN3_DOMU.201509250726Z) amd64
    # dmesg
    ...
    hypervisor0 at mainbus0: Xen version 4.2.amazon
    VIRQ_DEBUG interrupt using event channel 3
    vcpu0 at hypervisor0: Intel(R) Xeon(R) CPU E5-2680 v2 @ 2.80GHz, id 0x306e4
    vcpu1 at hypervisor0: Intel(R) Xeon(R) CPU E5-2680 v2 @ 2.80GHz, id 0x306e4 

1. The port to the NETWALKER (Cortex-A8) platform works as confirmed by Jun Ebihara, including instructions on how to set things up and dmesg output.

2. Ryota Ozaki is working on porting DTrace to ARM

3. Matt Thomas is making the ARM port ready to use multiple CPUs, see his posting, which shows a list of processes and their associated CPU.
   
   [Tags: arm, dmesg, dtrace, netwalker, smp]
   
   
   

Please join in and help test the code, and send your feedback to the lists. If no serious issues come up, the code will be merged within a week.

See Matt's posting to tech-kern for more details, inclusing diffs and links for amd64 and i386 GENERIC (+usbmp) kernels.

Further information on the state of the code - what is and what is not converted yet - can be found in the TODO.usbmp file.

[Tags: smp, usb]

``The NetBSD Foundation is pleased to announce completion of Multiprocessing Support for the port of its Open Source Operating System to the Xen hypervisor.

The NetBSD Fundation started the Xen MP project 8 month ago; the goal was to add SMP support to NetBSD/Xen domU kernels. This project has officially completed, and after a few bug fixes in the pmap(9) code it is now considered stable on both i386 and amd64. NetBSD 6.0 will ship with option MULTIPROCESSOR enabled by default for Xen domU kernels.

The availability of Xen MP support in NetBSD allows to run the NetBSD Open Source Operating Systems on a range of available infrastructure providers' systems. Amazon's Web Services with their Elastic Cloud Computing is a prominent examples here.

Xen is a virtualization software that enables several independent operating system instances ("domains") to run concurrently on the same computer hardware. The hardware is managed by the first domain (dom0), and further guest/user domains (domU) are spawned and managed by dom0. Operating systems available for running as dom0 and domU guests include Microsoft Windows, Solaris and Linux besides NetBSD.

NetBSD is a free, fast, secure, and highly portable Unix-like Open Source operating system. It is available for a wide range of platforms, from large-scale servers and powerful desktop systems to handheld and embedded devices. Its clean design and advanced features make it excellent for use in both production and research environments, and the source code is freely available under a business-friendly license. NetBSD is developed and supported by a large and vivid international community. Many applications are readily available through pkgsrc, the NetBSD Packages Collection.

NetBSD has been available for the Xen hypervisor since Xen 1 and NetBSD 2.0, released in 2004 , but until now only a single processor was supported in each NetBSD/xen domain.''

[Tags: amazon, ec2, smp, xen]

The article introduces real-time scheduling and the scheduling classes found in NetBSD 5.0, and gives an estimate on the response timeframe that can be expected for real-time applications. Setting scheduling policy and priority from a userland application is shown next, and programming examples for thread affinity, dynamic CPU sets and processor sets are shown. Besires C APIs, there are also a number or new commands in NetBSD 5.0 that can be used to control things from the command line, e.g. to define scheduling behaviour and manipulate processor sets. My favourite gem is the CPU used in the cpuctl(8) example, which is identified as "AMD Engineering Sample". :-)

[Tags: Articles, dmesg, posix, pthread, smp]

• Optimization for exec by using a cached copy of the file's exec header, and reducing locking-overhead by keeping locks instead of freeing them and the immediately re-locking them.

• The NetBSD kernel offers a few internal interfaces for allocating memory, depending on use of the memory, duration of use, size, etc. For some of the proper SMP-handling is important, and thus the number of CPUs has an impact on their performance. By adding caching, this can be mitigated to get linear behaviour, independent from the number of CPUs:

  

• The openat() and fstatat() system calls are not available in NetBSD yet, but other systems offer them, standards are about to pick them up, and ZFS assumes their presence. The system calls offer a way to way to specify a different directory to which relative paths are relative to, other than ".", by passing a file descriptor for that directory. Now there is a patch to add openat() and fstatat() to NetBSD.

• In theory, pipes are just a special case of (host-local) sockets, but it makes sense to keep a separate implementation for reasons of speed optimizations. NetBSD has the "PIPE_SOCKETPAIR" kernel option to force use of the socket code for pipes in order to reduce the memory footprint, but benchmarks reflect the performace hit. In order to improve the the situation, a number of improvements are under way, including better cache utilization and SMP-compliant memory allocation over homegrown memory management.

• Improved performance of exit(2) - this is important in environments with many short-running processes (think httpd, inetd).

• As a final step, freeing entries in the translation lookaside buffer (TLB) of the x86 (i386, amd64, xen) memory management unit AKA TLB shootdown were sped up to a point where TLB shootdown interrupts are 50% down during a system rebuild on an 8-cpu machine, and several million(!) calls page invalidation were optimized away, resulting in a 1% speed increase on the overall build.

• A partial(!) port of Sun's ZFS is also made available. It's not at the state where it can be used, but should be a good starting point for an experienced kernel hacker to continue working. See Andrew's mail to tech-kern for further directions.

SMP & audio: One area that hasn't been changed for moving towards fine-grained kernel locking was NetBSD's audio subsystem. As audio recording and playback is mostly done via interrupts, and as latency in those is critical, the audio subsystem was moved to the new interrupt handling system. The work can be found on the ad-audiomp branch, more information is available in Andrew's posting about the MP safe audio framework and drivers.

Benchmarking: Changing a system from inside out is a huge technical task. On the way, performance measurements and tuning are needed to make sure that the previous performance is still achieved while getting better performance in the desired development area. As a result, benchmarks results from Sun's libmicro benchmark suite were posted, which allow comparison not only against Linux and FreeBSD, but also between NetBSD-current and NetBSD 4.0, in order to identify if any bad effects were added. All performance tests were made on a machine with 8 CPUs, and the areas tested cover "small" (micro) areas like various system calls. Of course this doesn't lead to a 1:1 statement on how the systems will perform in a real-life scenario like e.g. in a database performance test, but it still help identifying problems and gives better hints where tuning can be done.

Another benchmark that was also made in that regard comes from Gregory McGarry, who has published performance measurements previously. This time, Gregory has run the lmbench 3.0 benchmark on recent NetBSD and FreeBSD systems as well as a number of previous NetBSD releases - useful for identifying performance degradation, too!

One other benchmark on dispatch latency run was made by Andrew Doran: on a machine that was (CPU-wise) loaded by some compile jobs, he started a two threads on a CPU that wasn't distracted by device interrupts, and measured how fast the scheduler reacted when one thread woke up the other one. The resulting graph shows that the scheduler handles the majority of requests in less than 10us - good enough for some realtime applications?

Kernel modules are another area that's under heavy change right now, and after recent changes to load modules from the bootloader and the kernel, the kernel build process was now changed so that pre-built kernel modules can be linked into a new kernel binary, resulting in a non-modular kernel. Eventually, this could mean that src/sys is built into separate modules, and that the (many) existing kernels that are present for each individual platform -- GENERIC, INSTALL is already gone, ALL, etc. etc. -- can be simply linked from pre-compiled modules, without recompiling things over again for each kernel. Of course the overal goal here is to speed up the system (and kernel!) build time, while maintaining maximum flexibility between modules and non-modular kernels.

With the progress in kernel modules, it is a question of time when the new kernel module handling supercedes the existing loadable kernel modules to such an extent that the latter will be completely removed from the system -- at least the latter was alredy proposed, but I'd prefer to see some documentation of the new system first. We'll see what comes first! (Documentation writers are always welcome! :-)

[Tags: benchmark, kernel, smp]

The story so far. Andrew Doran writes: `` The kernel synchronization model has been completely revamped since NetBSD 4.0, with the goal of making NetBSD a fully multithreaded, multiprocessor system with complete support for soft real-time applications.

Through NetBSD 4.0, the kernel used spinlocks and a per-CPU interrupt priority level (the spl(9) system) to provide mutual exclusion. These mechanisms did not lend themselves well to a multiprocessor environment supporting kernel preemption. Rules governing their use had been built up over many years of development, making them difficult to understand and use well. The use of thread based (lock) synchronization was limited and the available synchronization primitive (lockmgr) was inefficient and slow to execute.

In development branch that will becomple NetBSD 5.0, a new rich set of synchronization primitives and software tools have been developed to ease writing multithreaded kernel code that executes efficiently and safely on SMP systems. Some of these are:

• Thread-base adaptive mutexes. These are lightweight, exclusive locks that use threads as the focus of synchronization activity. Adaptive mutexes typically behave like spinlock, but under specific conditions an attempt to acquire an already held adaptive mutex may cause the acquring thread to sleep. Sleep activity occurs rarely. Busy-waiting is typically more efficient because mutex hold times are most often short. In contrast to pure spinlocks, a thread holding an adaptive mutex may be preempted in the kernel, which can allow for reduced latency where soft real-time application are in use on the system.

• Reader/writer locks. These are lightweight shared/exclusive locks that again use threads as the focus of synchronization activity. Their area of use is limited, most of it being in the file system code.

• CPU based spin mutexes, used mainly within the scheduler, device drivers and device driver support code. Pure spin mutexes are used when it is not safe, or impossible for, a thread to use a synchronization method that could block such as an adaptive mutex.

• Priority inheritance, implemented in support of soft real-time applications. Where a high priority thread is blocked waiting for a resource held by a lower priority thread, the kernel can temporarily "lend" a high priority level to the lower priority thread. This helps to ensure that the lower priority thread does no unduly obstruct the progress of the higher priority thread.

• Atomic operations. A full set of atomic operations implementing arithmetic and memory barrier operations has been provided. The atomic operations are available both in the kernel and to user applications, via the C library.

• Generic cross calls: a facility that allows one CPU or thread to easily make an arbitrary function call on any other CPUs in the system.

• The interrupt priority level interfaces (spl(9)) have long been used to block interrupts on a CPU-local basis. These interfaces have been simplified and streamlined to allow for code and algorithms that make use of low cost CPU-local synchronization. In addition, APIs are provided that allow detection of kernel preemption and allow the programmer to temporarily disable preemption across critical sections of code that cannot tolerate any interruption.

• "percpu" memory allocator: a simple memory allocator that provides arbitrary amounts of keyed storage. Allocations are replicated across all CPUs in the system and each CPU has its own private instance of any allocated object. Together, the cross call facility, atomic operations, spl(9) interfaces and percpu allocator make it easy to build lightweight, lock-free algorithms.

• Lockless memory allocators: the standard kernel memory allocators have been augmented with per-CPU caches which signficantly avoid costly synchronization overhead typically associated with allocation of memory on a multiprocessor system. ''

• New thread scheduler, named M2: it reduces the lock contention, and increases the thread affinity to avoid cache thrashing - this essentially improves the performance on SMP systems. M2 implements time-sharing class, and POSIX real-time extensions, used by soft real-time applications.

• Processor sets and affinity API provides possibility to bind the processes or threads to the specific CPU or group of CPUs. This allows applications to achieve better concurrency, CPU utilization, avoid cache thrashing and thus improve the performance on SMP systems.''

• A scheduler is responsible for distributing workdload on CPUs, and besides the 4BSD scheduler, a more modern "M2"-scheduler was recently added to NetBSD, see above. Parts of that scheduler were now suggested to be included in the general scheduling framework. That way, the 4BSD scheduler gets processor affinity (so threads / processes keep stuck to a single CPU and thus reduce cache misses when migrating between CPUs/cores).

  With other changes surrounding this, NetBSD-current beats FreeBSD 7.0 and all earlier NetBSD releases when running build.sh (i.e. compiling the whole system) on a 8-core machine. In the image, small values mean less time for the build, and are thus good. I find the results impressive. For more information, see Andrew's posting to tech-kern.

• Reader/writer locks are a locking primitive used to allow multiple readers, but to block them if one or more processes want to write to a ressource. Those locks are used in the NetBSD kernel, see the rwlock(9) manpage. In order to further optimize the performance of the rwlock primitives, a few optimizations were suggested by Andrew Doran which reduces the build time on an 8-cpu machine by 4%: ``There is less idle time during the build because the windows where a rwlock can be held but the owner is not running is reduced''.

• Another optimization was suggested which cuts down another 5% of the time for a complete system build via build.sh on an 8-core machine, this time by replacing a linear list of locks in the namei cache with a hash table for the locks. The namei cache helps to speed up translations from a path name to the corresponding vnodes, see namei(9).