[net] Measure DNS.CacheEvicted and DNS.CacheExpired.

[net] Measure DNS.CacheEvicted and DNS.CacheExpired.

On every eviction from HostCache record the time left to expiration
(DNS.CacheEvicted) or time since expiration (DNS.CacheExpired).
A distinction is made for evictions on Lookup (DNS.CacheExpiredOnGet).


Committed: http://src.chromium.org/viewvc/chrome?view=rev&revision=166781

[szym, 2012-11-03 21:58:15]

I'd like to be able to measure cache pressure before I start experimenting with
cache sizes / minimum TTL.

[cbentzel, 2012-11-05 14:47:20]

Why does this need to be a template argument, instead of a Callback or function
pointer?

[szym, 2012-11-05 15:37:47]

On 2012/11/05 14:47:20, cbentzel wrote:
> Why does this need to be a template argument, instead of a Callback or
function
> pointer?

It does not need to. I have a slight preference since Callback/function pointer
will add binary overhead (in machine code generated and processing), for limited
gain in readability:

template<...>
class ExpiringCache {
 ...
 public:
  typedef Callback<void(const KeyType& key,
                        const ValueType& value,
                        const ExpirationType& expiration,
                        const ExpirationType& now,
                        bool onGet)>) EvictionCallback;
  ExpiringCache(size_t max_entries, const EvictionCallback& eviction_callback)
    : max_entries_(max_entries), eviction_callback_(eviction_callback) {}

  ...
  void Evict(...) {
    if (!eviction_callback_.is_null())
      eviction_callback_.Run(...);
  }
};

The nice thing is that there would be no need to forward-declare
HostCache::EvictionHandler (could be in anonymous namespace, although callback.h
would have to be included in full I think), and it could actually hold state
(for more sophisticated experiments).

[jar (doing other things), 2012-11-05 17:44:56]

The use of the histograms LGTM.  

I hadn't read this code before, and was a bit surprised to see so much use of
templates (which I view as mostly write-only code).  After a bit of reading, I
could make sense of it... and assume the reason was to increase testability...
but still  I wasn't sure about the tradeoff.

I think we also try as hard as possible to have as little code in headers, so as
to expedite compilations.  I think this code tended to put a lot of code in
headers, and not all of it was seemingly that necessary (not sure).

Nits below.

https://codereview.chromium.org/11359053/diff/7001/net/base/expiring_cache.h
File net/base/expiring_cache.h (right):

https://codereview.chromium.org/11359053/diff/7001/net/base/expiring_cache.h#...
net/base/expiring_cache.h:47: //                 std::less<base::TimeTicks> >
cache(0);
This looks like it tries to create a cache with zero entries.  Reading the code,
it creates a cache with exactly 1 entry.

Is this just a bogus example, or were you trying to demonstrate something?

https://codereview.chromium.org/11359053/diff/7001/net/base/expiring_cache.h#...
net/base/expiring_cache.h:51: //   // Add another value that expires in 10
minutes.
Given the cache size of zero, this appear to replace, rather than add a new
value.

https://codereview.chromium.org/11359053/diff/7001/net/base/host_cache.cc
File net/base/host_cache.cc (right):

https://codereview.chromium.org/11359053/diff/7001/net/base/host_cache.cc#new...
net/base/host_cache.cc:104: base::TimeDelta::FromSeconds(1),
base::TimeDelta::FromDays(1), 100);
nit: early returns tend to make this sort of code easier to read.  Otherwise
readers have to scan forward to check on the logic in the else clauses, and the
remainder of the function.

[szym, 2012-11-05 18:28:44]

On 2012/11/05 17:44:56, jar wrote:
> The use of the histograms LGTM.  
> 
> I hadn't read this code before, and was a bit surprised to see so much use of
> templates (which I view as mostly write-only code).  After a bit of reading, I
> could make sense of it... and assume the reason was to increase testability...
> but still  I wasn't sure about the tradeoff.

My understanding of http://codereview.chromium.org/9407001 and
http://codereview.chromium.org/9436003 (by rsleevi) is that there's shared logic
between HostCache and CertVerifier (now MultiThreadedCertVerifier), and a
template (rather than virtual dispatch) is used for performance.
 
> I think we also try as hard as possible to have as little code in headers, so
as
> to expedite compilations.  I think this code tended to put a lot of code in
> headers, and not all of it was seemingly that necessary (not sure).
> 
> Nits below.
> 
> https://codereview.chromium.org/11359053/diff/7001/net/base/expiring_cache.h
> File net/base/expiring_cache.h (right):
> 
>
https://codereview.chromium.org/11359053/diff/7001/net/base/expiring_cache.h#...
> net/base/expiring_cache.h:47: //                 std::less<base::TimeTicks> >
> cache(0);
> This looks like it tries to create a cache with zero entries.  Reading the
code,
> it creates a cache with exactly 1 entry.
> 
> Is this just a bogus example, or were you trying to demonstrate something?

I did not write this, but I agree that Compact will fail to bring the size to
_below_ |max_entries_| so there's no way to create an ExpiringCache that does
not cache. This is not an issue for any other value of |max_entries_|. Not sure
if it matters in practice.

This implementation (Compact before insert) also prefers to keep a new entry
(even if expired) over any old entry (even if not expired). Not sure if it
matters in practice.
 
>
https://codereview.chromium.org/11359053/diff/7001/net/base/expiring_cache.h#...
> net/base/expiring_cache.h:51: //   // Add another value that expires in 10
> minutes.
> Given the cache size of zero, this appear to replace, rather than add a new
> value.
> 
> https://codereview.chromium.org/11359053/diff/7001/net/base/host_cache.cc
> File net/base/host_cache.cc (right):
> 
>
https://codereview.chromium.org/11359053/diff/7001/net/base/host_cache.cc#new...
> net/base/host_cache.cc:104: base::TimeDelta::FromSeconds(1),
> base::TimeDelta::FromDays(1), 100);
> nit: early returns tend to make this sort of code easier to read.  Otherwise
> readers have to scan forward to check on the logic in the else clauses, and
the
> remainder of the function.

Will fix.

[cbentzel, 2012-11-07 10:40:22]

LGTM, with nits.

It seems like the main thing you want to measure is what fraction of mutations
result in evictions of non-stale items, and I don't see that here. Or perhaps
ratio of items_evicted_due_to_size_limitations/all_items_evicted (the fraction
of mutations resulting in evictions may be bad data, because when it has reached
capacity it may only be replacing one each cycle since the current cache doesn't
try to go to a low-water mark of .8*capacity or something like that)

http://codereview.chromium.org/11359053/diff/7001/net/base/expiring_cache.h
File net/base/expiring_cache.h (right):

http://codereview.chromium.org/11359053/diff/7001/net/base/expiring_cache.h#n...
net/base/expiring_cache.h:22: void handle(const KeyType& key,
Nit: Should be Handle rather than handle.

http://codereview.chromium.org/11359053/diff/7001/net/base/expiring_cache.h#n...
net/base/expiring_cache.h:26: bool onGet) const {
Nit: on_get instead of onGet.

[szym, 2012-11-07 20:52:12]

On Wed, Nov 7, 2012 at 5:40 AM, <[email protected]> wrote:

> LGTM, with nits.
>
> It seems like the main thing you want to measure is what fraction of
> mutations
> result in evictions of non-stale items, and I don't see that here. Or
> perhaps
> ratio of items_evicted_due_to_size_**limitations/all_items_evicted (the
> fraction
> of mutations resulting in evictions may be bad data, because when it has
> reached
> capacity it may only be replacing one each cycle since the current cache
> doesn't
> try to go to a low-water mark of .8*capacity or something like that)
>

I'm not sure I understand your concern. Non-stale items are totaled by
sum(DNS.CacheEvicted), while stale items are totaled by
sum(DNS.CacheExpired). So your ratio could be evicted/(evicted+expired).


>
> http://codereview.chromium.**org/11359053/diff/7001/net/**
>
base/expiring_cache.h<http://codereview.chromium.org/11359053/diff/7001/net/base/expiring_cache.h>
> File net/base/expiring_cache.h (right):
>
> http://codereview.chromium.**org/11359053/diff/7001/net/**
>
base/expiring_cache.h#**newcode22<http://codereview.chromium.org/11359053/diff/7001/net/base/expiring_cache.h#newcode22>
> net/base/expiring_cache.h:22: void handle(const KeyType& key,
> Nit: Should be Handle rather than handle.
>
> http://codereview.chromium.**org/11359053/diff/7001/net/**
>
base/expiring_cache.h#**newcode26<http://codereview.chromium.org/11359053/diff/7001/net/base/expiring_cache.h#newcode26>
> net/base/expiring_cache.h:26: bool onGet) const {
> Nit: on_get instead of onGet.
>
>
http://codereview.chromium.**org/11359053/<http://codereview.chromium.org/113...
>

[commit-bot: I haz the power, 2012-11-08 18:47:13]

CQ is trying da patch. Follow status at
https://chromium-status.appspot.com/cq/[email protected]/11359053/11002

[commit-bot: I haz the power, 2012-11-08 18:47:14]

Failed to apply patch for net/base/host_cache.cc:
While running patch -p1 --forward --force --no-backup-if-mismatch;
  patching file net/base/host_cache.cc
  Hunk #1 FAILED at 5.
  Hunk #2 succeeded at 92 with fuzz 1 (offset 1 line).
  1 out of 2 hunks FAILED -- saving rejects to file net/base/host_cache.cc.rej

Patch:       net/base/host_cache.cc
Index: net/base/host_cache.cc
diff --git a/net/base/host_cache.cc b/net/base/host_cache.cc
index
5221fca9caeae744ad36b4b477f24afd0592dee4..cdd3008fbcb77a11f035366857cd99084d8413a4
100644
--- a/net/base/host_cache.cc
+++ b/net/base/host_cache.cc
@@ -5,6 +5,7 @@
 #include "net/base/host_cache.h"
 
 #include "base/logging.h"
+#include "base/metrics/histogram.h"
 #include "net/base/net_errors.h"
 
 namespace net {
@@ -91,4 +92,25 @@ scoped_ptr<HostCache> HostCache::CreateDefaultCache() {
   return make_scoped_ptr(new HostCache(kMaxHostCacheEntries));
 }
 
+void HostCache::EvictionHandler::Handle(
+    const Key& key,
+    const Entry& entry,
+    const base::TimeTicks& expiration,
+    const base::TimeTicks& now,
+    bool on_get) const {
+  if (on_get) {
+    DCHECK(now >= expiration);
+    UMA_HISTOGRAM_CUSTOM_TIMES("DNS.CacheExpiredOnGet", now - expiration,
+        base::TimeDelta::FromSeconds(1), base::TimeDelta::FromDays(1), 100);
+    return;
+  }
+  if (expiration > now) {
+    UMA_HISTOGRAM_CUSTOM_TIMES("DNS.CacheEvicted", expiration - now,
+        base::TimeDelta::FromSeconds(1), base::TimeDelta::FromDays(1), 100);
+  } else {
+    UMA_HISTOGRAM_CUSTOM_TIMES("DNS.CacheExpired", now - expiration,
+        base::TimeDelta::FromSeconds(1), base::TimeDelta::FromDays(1), 100);
+  }
+}
+
 }  // namespace net

[commit-bot: I haz the power, 2012-11-08 20:30:13]

CQ is trying da patch. Follow status at
https://chromium-status.appspot.com/cq/[email protected]/11359053/10002

[commit-bot: I haz the power, 2012-11-08 22:26:50]

Change committed as 166781