IEEE Consumer Electronics Magazine - May/June 2022 - 83

Figure 4. Updated message/packet header.
write-back, which usually are discarded
(dropped). The LLC bank controller (LLC CTLR)
never expects these blocks for write-back. They
are kept stored in the local routers to increase
the chances of local reply (to be explained in the
" Replying to Re-reference Requests from
Routers " section). Since the clean blocks, which
are in S state, have multiple sharers, only the
blocks in E state are sent for write-back to avoid
incoherence (to be explained in the " Maintaining
Cache Coherence " section). Clean write-back
messages also use the Store flag, much like dirty
write-back messages. To facilitate local reply,
both of these messages use another 1-bit flag
Dirty. The actions of the SRFD unit in the modified
routers are shown in Figure 3.
Replying to Re-Reference Requests From
Routers
The proposed architecture identifies read/
write requests in the local routers on their way
to the destination. Read and write requests are
identified using 1-bit flags S and X, respectively,
(refer Figure 4). When a new packet enters the
local router with its S flag SET (read request),
the requested address is compared with the
addresses of the stored blocks (packets). If SRFD
unit finds a match, it initiates a local reply to the
read request by swapping the source and destination
between the matched block (stored
packet) and the read request (new packet). SRFD
unit enables the VA and SA for the stored packet,
which were disabled during local store (refer
section " Storing Evicted L1 Cache Blocks in
Routers " ). The stored packet simply gets ejected
through the local output port as its destination
L1 cache is connected to the very same router.
Local reply reduces miss penalty significantly by
avoiding travel to fetch the requested block. In
essence, a read request is satisfied with a stored
clean/dirty block from the local router.
SRFD unit works similarly when a new
packet enters the local router with its X flag
SET (write request). However, there is an additional
check of the Dirty flag to identify whether
May/June 2022
the matched block is clean or dirty. Only when
the Dirty flag is found SET, a write request can
be satisfied with a local reply. If not SET
(matched block is clean), then the write request
travels toward LLC bank to fetch the requested
block (to be explained in the " Maintaining
Cache Coherence " section). The matched clean
block is dropped before the write request proceeds
toward LLC bank to avoid leaving an old
version of the block behind. Note that a write
request can be satisfied only with a matching
block, which is dirty and not clean.
Forwarding/Dropping of Stored Cache Blocks
When a new packet does not get a free buffer
(VC) in the local router for injection, SRFD unit
vacates a VC occupied by the stored blocks
(evicted clean or dirty). The oldest one is
vacated in case multiple VCs have stored blocks.
After the VC to be vacated is identified, SRFD
unit checks the Dirty flag of the block stored in
that VC. If the Dirty flag is found SET, the stored
block is dirty and, hence, is forwarded for writeback
by enabling VA and SA. Otherwise, the
stored block is clean and is silently dropped
since the LLC bank already has the block.
There can be other applications waiting in
the LLC CTLR to get the status of evicted L1
cache blocks to resume their execution. Hence,
even in the absence of injection pressure, when
the VCs are not yet full, some of the stored
blocks need to be either forwarded or dropped.
The proposed architecture employs two techniques
to identify the VC to be vacated: Time-Triggered
VC Vacate (TT-VCV) and MessageTriggered
VC Vacate (MT-VCV). TT-VCV dictates
that an evicted block remains stored in the local
router till the expiry of a specific time threshold.
With each VC, a threshold counter (ti)is
attached in the local input port (refer Figure 3).
MT-VCV dictates that when the status of a stored
block is requested by the LLC CTLR, the VC containing
the stored block is vacated.
Maintaining Cache Coherence
Evicted L1 blocks are stored only in the local
routers. Dirty blocks are always private (no
sharer), so when they are stored in local routers,
there is no violation of coherence. However,
when local replies are generated with these
83
IEEE Consumer Electronics Magazine - May/June 2022

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - May/June 2022
