Kernel Korner - Using RCU in the Linux 2.5 Kernel
The Linux hacker's toolbox already contains numerous symmetric multiprocessing (SMP) tools, so why bother with read-copy update (RCU)? Figure 1 answers this question, presenting hash-lookup performance with per-bucket locks on a four-CPU, 700MHz Pentium III system. Your mileage will vary with different workloads and on different hardware. For an excellent write-up on the use of other SMP techniques, see Robert Love's article in the August 2002 issue of Linux Journal [available at /article/5833].
All accesses are read-only, so one might expect rwlock to work as well as this system. However, one would be mistaken; rwlock actually scales negatively from one to two CPUs, partly because this variant of rwlock avoids starvation, thus incurring greater overhead. A much larger critical section is required for rwlock to be helpful. Although rwlock beats refcnt (a spinlock and reference counter) for small numbers of CPUs, even refcnt beats rwlock at four CPUs. In both cases, the scaling is atrocious; refcnt at four CPUs achieves only 54% of the ideal four-CPU performance, and rwlock achieves only 39%.
Simple spinlock incurs less overhead than either rwlock or refcnt, and it also scales somewhat better at 57%. But this scaling is still quite poor. Although some spinning occurs, due to CPUs attempting to access the same hash chain, such spinning accounts for less than one-quarter of the 43% degradation at four CPUs.
Only brlock scales linearly. However, brlock's single-CPU performance is subpar, requiring more than 300 nanoseconds to search a single-element hash chain with simple integer comparison. This process should not take much more than 100ns to complete.
Figure 2 illustrates the past quarter century's progress in hardware performance. The features that make the new kids (brats) so proud, however, are double-edged swords in SMP systems.
Unfortunately, many algorithms fail to take advantage of the brat's strengths, because they were developed back when the old man was in his prime. Unless you like slow, stately computing, you need to work with the brat.
The increase in CPU clock frequency has been astounding—where the old man might have been able to interfere with AM radio signals, the young brat might be able to synthesize them digitally. But memory speeds have not increased nearly as fast as CPU clock rates, so a single DRAM access can cost the brat up to a thousand instructions. Although the brat compensates for DRAM latency with large caches, these caches cannot help data bounced among CPUs. For example, when a given CPU acquires a lock, the lock has a 75% chance of being in another CPU's cache. The acquiring CPU stalls until the lock reaches its cache.
Cacheline bouncing explains much of the scaling shortfall in Figure 1, but it does not explain poor single-CPU performance. When there is only one CPU, no other caches are present in which the locks might hide. This is where the brat's 20-stage pipeline shows its dark side. SMP code must ensure that no critical section's instructions or memory operations bleed out into surrounding code. After all, the whole point of a lock is to prevent multiple CPUs from concurrently executing any of the critical section's operations.
Memory barriers prevent such bleeding. These memory barriers are included implicitly in atomic instructions on x86 CPUs, but they are separate instructions on most other CPUs. In either case, locking primitives must include memory barriers. But these barriers cause pipeline flushes and stalls, the overhead of which increases with pipeline length. This overhead is responsible for the single-CPU slowness shown in Figure 1.
Table 1 outlines the costs of basic operations on 700MHz Intel Pentium III machines, which can retire two integer instructions per clock. The atomic operation timings assume the data already resides in the CPU's cache. All of these timings can vary, depending on the cache state, bus loading and the exact sequence of operations.
Table 1. Time Required for Common Operations on a 700MHz Pentium III
|L2 cache hit||12.9|
|cmpxchg atomic increment||107.3|
The overheads increase relative to instruction execution overhead. For example, on a 1.8GHz Pentium 4, atomic increment costs about 75ns—slower than the 700MHz Pentium III, despite having a more than twice as fast clock.
These overheads also explain rwlock's poor performance. The read-side critical section must contain hundreds of instructions for it to continue executing once some other CPU read acquires the lock, as illustrated in Figure 3. In this figure, the vertical arrows represent time passing on two pairs of CPUs, one pair using rwlock and the other using spinlock. The diagonal arrows represent data moving between the CPUs' caches. The rwlock critical sections do not overlap at all; the overhead of moving the lock from one CPU to the other rivals that of the critical section.
- On Your Marks, Get Set...Gutsy Gibbon!
- VMware's Clarity Design System
- Let's Go to Mars with Martian Lander
- Applied Expert Systems, Inc.'s CleverView for TCP/IP on Linux
- Papa's Got a Brand New NAS
- My Childhood in a Cigar Box
- Returning Values from Bash Functions
- Simplenote, Simply Awesome!
- Raspberry Pi: the Perfect Home Server
- GENIVI Alliance's GENIVI Vehicle Simulator