On 21/01/2013 12:10 a.m., Ralf Hildebrandt wrote:
I have a general question regarding Squid 3.2.x and the hot object
cache.
According to what I found in the FAQ, the "hot object cache" is
located in RAM, taking an amount of RAM limited by cache_mem.
So, frequently requested objects are stored in this memory area. But
what is a hot item?
An item which was requested very recently (temporally close to the
active traffic), with level of "heat" being the HIT rate on it. As has
been documented by the WikiMedia guys, you can get 100% TCP_MEM_HIT rate
when being "DoS'd" by genuine client requests for a popular object.
How big can a hot item be?
Up to the configured maximum_object_size_in_memory (default of 512 KB).
Are objects transferred
from the cache_mem to the disk cache (and vice versa).
Yes.
* New objects (which are small enough) start off in memory-only and when
they reach "cold" enough HIT rate they get swapped out to a disk store.
- object too big for memory get swapped out to disk immediately and
only a short windows of their size being actively used by a client is
kept in memory.
* In UFS/AUFS/diskd storage, objects stay on disk. When HIT thay can get
swapped back into memory as a duplicate (but not swapped back out again
when they go cold, unless invalidated by something).
* In COSS or Rock storage, object swapping in/out happens the same way.
But only to "pages" of the storage area which are in memory at the time.
Identical to a RAM-disk storage highly optimized to Squid access
patterns and the HTTP average object size ranges.
The reason for my question: WOuld a RAM disk help to speed up the proxy?
That is really complicated.
It would speed up if you had a cache configuration with access latency
worse than the RAM cache overhead costs.
For example; RAM-disk is clearly faster than SSD or disk for the old
popular UFS/AUFS/diskd storage type.
Rock and COSS storage types however are far more optimized for speed,
using both disk and RAM storage in ther normal "disk" configuration. Ao
a % of accesses from them will be *faster* than RAM-disk, and some will
be slower to a HDD than UFS. Add an SSD behind Rock or COSS and the
latency on those worst-case disk loads goes down to something comparible
to RAM-disk, but you face SSD design limitations on write-ops and
lifetime which are consumed by Squid faster than normal usage (ie they
die a bit faster than manufacturer specs would indicate - still usable
though).
Also, HTTP header paring is a significant overhead in the whole request
process and a sizeable % of the traffic bytes. By the time you are
getting close to RAM speeds on a large cache you are hitting traffic
rates where the parsing speed bottlenecks more than the disk I/O. With
HTTP/1.1 optimizations generating a lot of 304 responses the portion of
parsing goes up significantly.
So,... YMMV, if you wish give it a try. And if you find it actually *is*
faster some guys on the squid-dev mailing list would like to know by how
much and why.
Amos
