f8d9ccf8d5
5 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
Markus Armbruster
|
2a6a4076e1 |
Clean up ill-advised or unusual header guards
Cleaned up with scripts/clean-header-guards.pl. Signed-off-by: Markus Armbruster <armbru@redhat.com> Reviewed-by: Richard Henderson <rth@twiddle.net> |
||
|
Emilio G. Cota
|
909eaac9bb |
tb hash: track translated blocks with qht
Having a fixed-size hash table for keeping track of all translation blocks
is suboptimal: some workloads are just too big or too small to get maximum
performance from the hash table. The MRU promotion policy helps improve
performance when the hash table is a little undersized, but it cannot
make up for severely undersized hash tables.
Furthermore, frequent MRU promotions result in writes that are a scalability
bottleneck. For scalability, lookups should only perform reads, not writes.
This is not a big deal for now, but it will become one once MTTCG matures.
The appended fixes these issues by using qht as the implementation of
the TB hash table. This solution is superior to other alternatives considered,
namely:
- master: implementation in QEMU before this patchset
- xxhash: before this patch, i.e. fixed buckets + xxhash hashing + MRU.
- xxhash-rcu: fixed buckets + xxhash + RCU list + MRU.
MRU is implemented here by adding an intermediate struct
that contains the u32 hash and a pointer to the TB; this
allows us, on an MRU promotion, to copy said struct (that is not
at the head), and put this new copy at the head. After a grace
period, the original non-head struct can be eliminated, and
after another grace period, freed.
- qht-fixed-nomru: fixed buckets + xxhash + qht without auto-resize +
no MRU for lookups; MRU for inserts.
The appended solution is the following:
- qht-dyn-nomru: dynamic number of buckets + xxhash + qht w/ auto-resize +
no MRU for lookups; MRU for inserts.
The plots below compare the considered solutions. The Y axis shows the
boot time (in seconds) of a debian jessie image with arm-softmmu; the X axis
sweeps the number of buckets (or initial number of buckets for qht-autoresize).
The plots in PNG format (and with errorbars) can be seen here:
http://imgur.com/a/Awgnq
Each test runs 5 times, and the entire QEMU process is pinned to a
single core for repeatability of results.
Host: Intel Xeon E5-2690
28 ++------------+-------------+-------------+-------------+------------++
A***** + + + master **A*** +
27 ++ * xxhash ##B###++
| A******A****** xxhash-rcu $$C$$$ |
26 C$$ A******A****** qht-fixed-nomru*%%D%%%++
D%%$$ A******A******A*qht-dyn-mru A*E****A
25 ++ %%$$ qht-dyn-nomru &&F&&&++
B#####% |
24 ++ #C$$$$$ ++
| B### $ |
| ## C$$$$$$ |
23 ++ # C$$$$$$ ++
| B###### C$$$$$$ %%%D
22 ++ %B###### C$$$$$$C$$$$$$C$$$$$$C$$$$$$C$$$$$$C
| D%%%%%%B###### @E@@@@@@ %%%D%%%@@@E@@@@@@E
21 E@@@@@@E@@@@@@F&&&@@@E@@@&&&D%%%%%%B######B######B######B######B######B
+ E@@@ F&&& + E@ + F&&& + +
20 ++------------+-------------+-------------+-------------+------------++
14 16 18 20 22 24
log2 number of buckets
Host: Intel i7-4790K
14.5 ++------------+------------+-------------+------------+------------++
A** + + + master **A*** +
14 ++ ** xxhash ##B###++
13.5 ++ ** xxhash-rcu $$C$$$++
| qht-fixed-nomru %%D%%% |
13 ++ A****** qht-dyn-mru @@E@@@++
| A*****A******A****** qht-dyn-nomru &&F&&& |
12.5 C$$ A******A******A*****A****** ***A
12 ++ $$ A*** ++
D%%% $$ |
11.5 ++ %% ++
B### %C$$$$$$ |
11 ++ ## D%%%%% C$$$$$ ++
| # % C$$$$$$ |
10.5 F&&&&&&B######D%%%%% C$$$$$$C$$$$$$C$$$$$$C$$$$$C$$$$$$ $$$C
10 E@@@@@@E@@@@@@B#####B######B######E@@@@@@E@@@%%%D%%%%%D%%%###B######B
+ F&& D%%%%%%B######B######B#####B###@@@D%%% +
9.5 ++------------+------------+-------------+------------+------------++
14 16 18 20 22 24
log2 number of buckets
Note that the original point before this patch series is X=15 for "master";
the little sensitivity to the increased number of buckets is due to the
poor hashing function in master.
xxhash-rcu has significant overhead due to the constant churn of allocating
and deallocating intermediate structs for implementing MRU. An alternative
would be do consider failed lookups as "maybe not there", and then
acquire the external lock (tb_lock in this case) to really confirm that
there was indeed a failed lookup. This, however, would not be enough
to implement dynamic resizing--this is more complex: see
"Resizable, Scalable, Concurrent Hash Tables via Relativistic
Programming" by Triplett, McKenney and Walpole. This solution was
discarded due to the very coarse RCU read critical sections that we have
in MTTCG; resizing requires waiting for readers after every pointer update,
and resizes require many pointer updates, so this would quickly become
prohibitive.
qht-fixed-nomru shows that MRU promotion is advisable for undersized
hash tables.
However, qht-dyn-mru shows that MRU promotion is not important if the
hash table is properly sized: there is virtually no difference in
performance between qht-dyn-nomru and qht-dyn-mru.
Before this patch, we're at X=15 on "xxhash"; after this patch, we're at
X=15 @ qht-dyn-nomru. This patch thus matches the best performance that we
can achieve with optimum sizing of the hash table, while keeping the hash
table scalable for readers.
The improvement we get before and after this patch for booting debian jessie
with arm-softmmu is:
- Intel Xeon E5-2690: 10.5% less time
- Intel i7-4790K: 5.2% less time
We could get this same improvement _for this particular workload_ by
statically increasing the size of the hash table. But this would hurt
workloads that do not need a large hash table. The dynamic (upward)
resizing allows us to start small and enlarge the hash table as needed.
A quick note on downsizing: the table is resized back to 2**15 buckets
on every tb_flush; this makes sense because it is not guaranteed that the
table will reach the same number of TBs later on (e.g. most bootup code is
thrown away after boot); it makes sense to grow the hash table as
more code blocks are translated. This also avoids the complication of
having to build downsizing hysteresis logic into qht.
Reviewed-by: Sergey Fedorov <serge.fedorov@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Emilio G. Cota <cota@braap.org>
Message-Id: <1465412133-3029-15-git-send-email-cota@braap.org>
Signed-off-by: Richard Henderson <rth@twiddle.net>
|
||
|
Emilio G. Cota
|
42bd32287f |
tb hash: hash phys_pc, pc, and flags with xxhash
For some workloads such as arm bootup, tb_phys_hash is performance-critical. The is due to the high frequency of accesses to the hash table, originated by (frequent) TLB flushes that wipe out the cpu-private tb_jmp_cache's. More info: https://lists.nongnu.org/archive/html/qemu-devel/2016-03/msg05098.html To dig further into this I modified an arm image booting debian jessie to immediately shut down after boot. Analysis revealed that quite a bit of time is unnecessarily spent in tb_phys_hash: the cause is poor hashing that results in very uneven loading of chains in the hash table's buckets; the longest observed chain had ~550 elements. The appended addresses this with two changes: 1) Use xxhash as the hash table's hash function. xxhash is a fast, high-quality hashing function. 2) Feed the hashing function with not just tb_phys, but also pc and flags. This improves performance over using just tb_phys for hashing, since that resulted in some hash buckets having many TB's, while others getting very few; with these changes, the longest observed chain on a single hash bucket is brought down from ~550 to ~40. Tests show that the other element checked for in tb_find_physical, cs_base, is always a match when tb_phys+pc+flags are a match, so hashing cs_base is wasteful. It could be that this is an ARM-only thing, though. UPDATE: On Tue, Apr 05, 2016 at 08:41:43 -0700, Richard Henderson wrote: > The cs_base field is only used by i386 (in 16-bit modes), and sparc (for a TB > consisting of only a delay slot). > It may well still turn out to be reasonable to ignore cs_base for hashing. BTW, after this change the hash table should not be called "tb_hash_phys" anymore; this is addressed later in this series. This change gives consistent bootup time improvements. I tested two host machines: - Intel Xeon E5-2690: 11.6% less time - Intel i7-4790K: 19.2% less time Increasing the number of hash buckets yields further improvements. However, using a larger, fixed number of buckets can degrade performance for other workloads that do not translate as many blocks (600K+ for debian-jessie arm bootup). This is dealt with later in this series. Reviewed-by: Sergey Fedorov <sergey.fedorov@linaro.org> Reviewed-by: Richard Henderson <rth@twiddle.net> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Emilio G. Cota <cota@braap.org> Message-Id: <1465412133-3029-8-git-send-email-cota@braap.org> Signed-off-by: Richard Henderson <rth@twiddle.net> |
||
|
Peter Crosthwaite
|
41da4bd642 |
cpu-defs: Move out TB_JMP defines
These are not Architecture specific in any way so move them out of cpu-defs.h. tb-hash.h is an appropriate place as a leading user and their strong relationship to TB hashing and caching. Reviewed-by: Richard Henderson <rth@redhat.com> Signed-off-by: Peter Crosthwaite <crosthwaite.peter@gmail.com> Message-Id: <43ceca65a3fa240efac49aa0bf604ad0442e1710.1433052532.git.crosthwaite.peter@gmail.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> |
||
|
Peter Crosthwaite
|
e1b89321ba |
include/exec: Move tb hash functions out
This is one of very few things in exec-all with a genuine CPU architecture dependency. Move these hashing helpers to a new header to trim exec-all.h down to a near architecture-agnostic header. The defs are only used by cpu-exec and translate-all which are both arch-obj's so the new tb-hash.h has no core code usage. Reviewed-by: Richard Henderson <rth@redhat.com> Signed-off-by: Peter Crosthwaite <crosthwaite.peter@gmail.com> Message-Id: <9d048b96f7cfa64a4d9c0b88e0dd2877fac51d41.1433052532.git.crosthwaite.peter@gmail.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> |